U.S. patent number 10,249,016 [Application Number 15/868,694] was granted by the patent office on 2019-04-02 for graphics processing units and methods using cost indications for sets of tiles of a rendering space.
This patent grant is currently assigned to Imagination Technologies Limited. The grantee listed for this patent is Imagination Technologies Limited. Invention is credited to Richard Broadhurst, Steven Fishwick, John W. Howson.
![](/patent/grant/10249016/US10249016-20190402-D00000.png)
![](/patent/grant/10249016/US10249016-20190402-D00001.png)
![](/patent/grant/10249016/US10249016-20190402-D00002.png)
![](/patent/grant/10249016/US10249016-20190402-D00003.png)
![](/patent/grant/10249016/US10249016-20190402-D00004.png)
![](/patent/grant/10249016/US10249016-20190402-D00005.png)
![](/patent/grant/10249016/US10249016-20190402-D00006.png)
![](/patent/grant/10249016/US10249016-20190402-D00007.png)
![](/patent/grant/10249016/US10249016-20190402-D00008.png)
![](/patent/grant/10249016/US10249016-20190402-D00009.png)
![](/patent/grant/10249016/US10249016-20190402-D00010.png)
View All Diagrams
United States Patent |
10,249,016 |
Broadhurst , et al. |
April 2, 2019 |
Graphics processing units and methods using cost indications for
sets of tiles of a rendering space
Abstract
A graphics processing unit is configured to process graphics
data using a rendering space which is sub-divided into a plurality
of tiles. The graphics processing unit comprises one or more
processing cores configured to process graphics data. The graphics
processing unit also comprises cost indication logic configured to
obtain a cost indication for each of a plurality of sets of one or
more tiles of the rendering space, wherein the cost indication for
a set of one or more tiles is suggestive of a cost of processing
the set of one or more tiles. The graphics processing unit also
comprises scheduling logic configured to schedule, in dependence
upon the cost indications, the sets of one or more tiles for
processing on the one or more processing cores.
Inventors: |
Broadhurst; Richard (Kings
Langley, GB), Howson; John W. (Kings Langley,
GB), Fishwick; Steven (Kings Langley, GB) |
Applicant: |
Name |
City |
State |
Country |
Type |
Imagination Technologies Limited |
Kings Langley |
N/A |
GB |
|
|
Assignee: |
Imagination Technologies
Limited (Kings Langley, GB)
|
Family
ID: |
58463385 |
Appl.
No.: |
15/868,694 |
Filed: |
January 11, 2018 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20180197271 A1 |
Jul 12, 2018 |
|
Foreign Application Priority Data
|
|
|
|
|
Jan 12, 2017 [GB] |
|
|
1700562.0 |
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06T
1/20 (20130101); G09G 5/363 (20130101); G06T
15/005 (20130101); G06T 2210/52 (20130101) |
Current International
Class: |
G06T
1/20 (20060101); G06T 15/00 (20110101); G09G
5/36 (20060101) |
References Cited
[Referenced By]
U.S. Patent Documents
Foreign Patent Documents
|
|
|
|
|
|
|
105261066 |
|
Jan 2016 |
|
CN |
|
2481099 |
|
Dec 2011 |
|
GB |
|
2013/178245 |
|
Dec 2013 |
|
WO |
|
Primary Examiner: Richer; Joni
Attorney, Agent or Firm: Vorys, Sater, Seymour and Pease LLP
DeLuca; Vincent M
Claims
What is claimed is:
1. A graphics processing unit configured to process graphics data
using a rendering space which is sub-divided into a plurality of
tiles, the graphics processing unit comprising: a plurality of
processing cores configured to render graphics data; cost
indication logic configured to obtain a cost indication for each of
a plurality of sets of one or more tiles of the rendering space,
wherein the cost indication for a set of one or more tiles is
suggestive of a cost of processing the set of one or more tiles,
wherein the cost indication logic is configured to quantise the
cost indications to be represented by a number of bits, the number
of bits being in a range from 1 to 8, and wherein the graphics
processing unit is configured to store the cost indications; and
scheduling logic configured to assign, in dependence upon the cost
indications, the sets of one or more tiles to the processing cores
for rendering.
2. The graphics processing unit of claim 1 wherein the scheduling
logic assigns the sets of one or more tiles for a render to the
processing cores in an order depending upon the cost
indications.
3. The graphics processing unit of claim 1 wherein the scheduling
logic assigns a plurality of sets with the highest cost indications
to different processing cores.
4. The graphics processing unit of claim 1 wherein the cost
indication logic is configured to determine the cost indications
for the sets of one or more tiles of the rendering space.
5. The graphics processing unit of claim 4 wherein the cost
indication logic is configured to determine a cost indication for a
tile of the rendering space by combining scores associated with
primitives which are present in the tile, wherein the score
associated with a primitive is dependent upon an object type of an
object of which the primitive is a part.
6. The graphics processing unit of claim 4 wherein the cost
indication logic is configured to determine a cost indication for a
tile of the rendering space based on one or more of the following
factors: (i) a number of primitives in the tile; (ii) object types
associated with the primitives in the tile; (iii) tile coverage
area of the primitives in the tile; (iv) characteristics of one or
more shader programs which are to be executed for rendering the
primitives in the tile; (v) a user input; and (vi) a processing
cost of a corresponding tile in a previous render.
7. The graphics processing unit of claim 6 wherein the
characteristics of a shader program include one or more of: (i) a
length of the shader program; (ii) an amount of resources or
registers used by the shader program; (iii) whether the shader
program includes conditional flow control; (iv) whether the shader
program includes loops for which the number of repetitions is
undefined at compile time; and (v) an amount of memory reads and/or
writes used in the shader program.
8. The graphics processing unit of claim 6 wherein the graphics
processing unit is configured to receive, from a compiler, the
characteristics of the one or more shader programs which are to be
executed for rendering the primitives in the tile.
9. The graphics processing unit of claim 1 wherein the scheduling
logic is configured to form a priority queue for sets of tiles for
each of the different quantised cost indication values, and to
assign the sets of one or more tiles to the processing cores for
rendering in an order according to the contents of the priority
queues.
10. The graphics processing unit of claim 1 wherein the sets of one
or more tiles each comprise a plurality of tiles.
11. The graphics processing unit of claim 10 wherein the cost
indication logic is configured to: obtain a respective cost
indication for each of the tiles of a set of tiles; and determine
the cost indication for the set of tiles based on the cost
indications of the tiles within that set.
12. The graphics processing unit of claim 11 wherein the cost
indication logic is configured to determine the cost indication for
a set of tiles by determining: (i) an average of the cost
indications of the tiles within that set, (ii) a sum of the cost
indications of the tiles within that set, (iii) a maximum of the
cost indications of the tiles within that set, or (iv) a number of
the tiles within that set which have cost indications above a
threshold.
13. The graphics processing unit of claim 10 wherein the scheduling
logic is configured to determine a processing order of tiles within
a particular set of tiles based on the cost indications of the
tiles within that particular set.
14. The graphics processing unit of claim 1 wherein the graphics
processing unit comprises geometry processing logic and rendering
logic, wherein the geometry processing logic includes a tiling unit
configured to generate control streams for the tiles of the
rendering space indicating which primitives are present in the
tiles, wherein the cost indication logic is implemented as part of
the geometry processing logic, wherein the rendering logic
comprises the processing cores and is configured to render
primitives in tiles of the rendering space in accordance with the
generated control streams, and wherein the scheduling logic
determines the assignment of sets of one or more tiles to the
processing cores of the rendering logic.
15. The graphics processing unit of claim 14 wherein the tiling
unit is configured to include the cost indications in the control
streams for the tiles.
16. The graphics processing unit of claim 14 wherein the scheduling
logic is configured to add sets of tiles to respective rendering
queues for the processing cores during rendering in accordance with
the occupancy of the rendering queues.
17. A method of processing graphics data in a graphics processing
system which comprises a plurality of processing cores configured
to render graphics data, the graphics processing system being
configured to use a rendering space which is sub-divided into a
plurality of tiles, the method comprising: obtaining a cost
indication for each of a plurality of sets of one or more tiles of
the rendering space, wherein the cost indication for a set of one
or more tiles is suggestive of a cost of processing the set of one
or more tiles, wherein the cost indications are quantised to be
represented by a number of bits, the number of bits being in a
range from 1 to 8; storing the cost indications; and assigning, in
dependence upon the cost indications, the sets of one or more tiles
to the processing cores for rendering.
18. The method of claim 17 further comprising forming a priority
queue for sets of tiles for each of the different quantised cost
indication values, wherein the sets of one or more tiles are
assigned to the processing cores for rendering in an order
according to the contents of the priority queues.
19. A non-transitory computer readable storage medium having stored
thereon a computer readable description of an integrated circuit
that, when processed in an integrated circuit manufacturing system,
causes the integrated circuit manufacturing system to manufacture a
graphics processing unit which is configured to process graphics
data using a rendering space which is sub-divided into a plurality
of tiles, wherein the graphics processing unit comprises: a
plurality of processing cores configured to render graphics data;
cost indication logic configured to obtain a cost indication for
each of a plurality of sets of one or more tiles of the rendering
space, wherein the cost indication for a set of one or more tiles
is suggestive of a cost of processing the set of one or more tiles,
wherein the cost indication logic is configured to quantise the
cost indications to be represented by a number of bits, the number
of bits being in a range from 1 to 8, and wherein the graphics
processing unit is configured to store the cost indications; and
scheduling logic configured to assign, in dependence upon the cost
indications, the sets of one or more tiles to the processing cores
for rendering.
Description
BACKGROUND
Graphics processing systems are typically configured to receive
graphics data, e.g. from an application running on a computer
system, and to render the graphics data to provide a rendering
output. For example, the graphics data provided to a graphics
processing system may describe geometry within a three dimensional
(3D) scene to be rendered, and the rendering output may be a
rendered image of the scene. Some graphics processing systems
(which may be referred to as "tile-based" graphics processing
systems) use a rendering space which is subdivided into a plurality
of tiles. The "tiles" are regions of the rendering space, and may
have any suitable shape, but are typically rectangular (where the
term "rectangular" includes square). To give some examples, a tile
may cover a 16.times.16 block of pixels or a 32.times.32 block of
pixels of an image to be rendered. As is known in the art, there
are many benefits to subdividing the rendering space into tiles.
For example, subdividing the rendering space into tiles allows an
image to be rendered in a tile-by-tile manner, wherein graphics
data for a tile can be temporarily stored "on-chip" during the
rendering of the tile.
Tile-based graphics processing systems typically operate in two
phases: a geometry processing phase and a rendering phase. In the
geometry processing phase, the graphics data for a render is
analysed to determine, for each of the tiles, which graphics data
items are present within that tile. Then in the rendering phase, a
tile can be rendered by processing those graphics data items which
are determined to be present within that tile (without needing to
process graphics data items which were determined in the geometry
processing phase to not be present within the particular tile). The
graphics data items may represent geometric shapes, which describe
surfaces of structures in the scene, and which are referred to as
"primitives". A common primitive shape is a triangle, but
primitives may be other 2D shapes or may be lines or points also.
Objects can be composed of one or more (e.g. hundreds, thousands or
millions) of such primitives.
FIG. 1 shows some elements of a graphics processing system 100
which may be used to render an image of a 3D scene. The graphics
processing system 100 comprises a graphics processing unit (GPU)
102 and two portions of memory 104.sub.1 and 104.sub.2. The two
portions of memory 104.sub.1 and 104.sub.2 may, or may not, be
parts of the same physical memory.
The GPU 102 comprises a pre-processing module 106, a tiling unit
108 and rendering logic 110, wherein the rendering logic 110
comprises a fetch unit 112 and processing logic 113 which includes
one or more processing cores 114. The rendering logic 110 is
configured to use the processing cores 114 to implement hidden
surface removal (HSR) and texturing and/or shading on graphics data
(e.g. primitive fragments) for tiles of the rendering space.
The graphics processing system 100 is arranged such that a sequence
of primitives provided by an application is received at the
pre-processing module 106. In a geometry processing phase, the
pre-processing module 106 performs functions such as geometry
processing including clipping and culling to remove primitives
which do not fall into a visible view. The pre-processing module
106 may also project the primitives into screen-space. The
primitives which are output from the pre-processing module 106 are
passed to the tiling unit 108 which determines which primitives are
present within each of the tiles of the rendering space of the
graphics processing system 100. The tiling unit 108 assigns
primitives to tiles of the rendering space by creating control
streams (or "display lists") for the tiles, wherein the control
stream for a tile includes indications of primitives which are
present within the tile. The control streams and the primitives are
outputted from the tiling unit 108 and stored in the memory
104.sub.1.
In a rendering phase, the rendering logic 110 renders graphics data
for tiles of the rendering space to generate values of a render,
e.g. rendered image values. The rendering logic 110 may be
configured to implement any suitable rendering technique, such as
rasterisation or ray tracing to perform the rendering. In order to
render a tile, the fetch unit 112 fetches the control stream for a
tile and the primitives relevant to that tile from the memory
104.sub.1. For example, the rendering unit may implement
rasterisation according to a deferred rendering technique, such
that one or more of the processing core(s) 114 are used to perform
hidden surface removal to thereby remove fragments of primitives
which are hidden in the scene, and then one or more of the
processing core(s) 114 are used to apply texturing and/or shading
to the remaining primitive fragments to thereby form rendered image
values. Methods of performing hidden surface removal and
texturing/shading are known in the art. The term "fragment" refers
to a sample of a primitive at a sampling point, which is to be
processed for rendering pixels of an image. In some examples, there
may be a one to one mapping of sample positions to pixels. In other
examples there may be more sample positions than pixels, and this
oversampling can allow for higher quality rendering of pixel
values, e.g. by facilitating anti-aliasing and other filtering that
may be applied to multiple fragments for rendering each of the
pixel values. The texturing and/or shading performed on the
fragments which pass the HSR stage determines pixel colour values
of a rendered image which can be passed to the memory 104.sub.2 for
storage in a frame buffer. Texture data may be received at the
rendering logic 110 from the memory 104.sub.1 in order to apply
texturing to the primitive fragments, as is known in the art.
Shader programs may be executed to apply shading to the primitive
fragments. The texturing/shading process may include applying
further processing to the primitive fragments (e.g. alpha blending
and other processes), as is known in the art in order to determine
rendered pixel values of an image. The rendering logic 110
processes primitives in each of the tiles and when the whole image
has been rendered and stored in the memory 104.sub.2, the rendered
image can be outputted from the graphics processing system 100 and
used in any suitable manner, e.g. displayed on a display or stored
in memory or transmitted to another device, etc.
In some systems, a particular processing core can be used to
perform hidden surface removal at one point in time and
texturing/shading at another point in time. In some other systems,
some of the processing cores are dedicated for performing hidden
surface removal whilst others of the processing cores are dedicated
for performing texturing and/or shading on primitive fragments.
The graphics processing system 100 described above is a deferred
rendering system because the rendering logic 110 is configured to
perform the HSR processing on a primitive fragment before the
texturing/shading processing is applied to the primitive fragment.
Other graphics processing systems are not deferred rendering system
in the sense that they are configured to perform the texturing
and/or shading of primitive fragments before the HSR is performed
on those primitive fragments. Deferred rendering systems avoid the
processing involved in applying texturing and/or shading to at
least some of the primitive fragments which are removed by the
hidden surface removal process.
If the rendering logic 110 includes more than one processing core
114 then the processing cores can process different data in
parallel, thereby improving the efficiency of the rendering logic
110. In some systems, the tiles are assigned to processing cores of
the rendering logic 110, such that the graphics data for rendering
a particular tile is processed in a single processing core. The
graphics data for rendering a different tile may be processed by a
different, single processing core. Processing a particular tile on
a single processing core (rather than spreading the processing of
the particular tile across multiple cores) can have benefits such
as an improved cache hit rate. Multiple tiles may be assigned to
the same processing core, which can be referred to as having
"multiple tiles in flight". When all of the tiles for a render have
been processed by the rendering logic 110, the render is complete.
Then the results of the render (e.g. a rendered frame) can be used
as appropriate (e.g. displayed on a display or stored in a memory
or transmitted to another device, etc.), and the rendering logic
110 can process tiles of a subsequent render.
SUMMARY
This Summary is provided to introduce a selection of concepts in a
simplified form that are further described below in the Detailed
Description. This Summary is not intended to identify key features
or essential features of the claimed subject matter, nor is it
intended to be used to limit the scope of the claimed subject
matter.
A graphics processing unit is configured to process graphics data
using a rendering space which is sub-divided into a plurality of
tiles. The graphics processing unit comprises one or more
processing cores configured to process graphics data. The graphics
processing unit may also comprise cost indication logic configured
to obtain a cost indication for each of a plurality of sets of one
or more tiles of the rendering space, wherein the cost indication
for a set of one or more tiles is suggestive of a cost of
processing the set of one or more tiles. The graphics processing
unit also comprises scheduling logic which may be configured to
schedule, in dependence upon the cost indications, the sets of one
or more tiles for processing on the one or more processing cores.
For example, in this way, sets of one or more tiles are scheduled
for processing on the processing core(s) according to how much work
is likely to be involved in the processing of those tiles. For
example, tiles which are likely to involve a lot of work can be
scheduled for processing before tiles which are likely to involve
less work. This can improve the efficiency of the graphics
processing system, e.g. in terms of the amount of time taken to
process the graphics data for all of the tiles in the rendering
space.
In some examples, the scheduling logic may be configured to
subdivide at least one set of one or more tiles of the rendering
space to form a plurality of subunits (e.g. subtiles) and to assign
at least some of those subunits to different processing cores for
rendering. For example, the scheduling logic may be configured to
subdivide at least one of the tiles of the rendering space to form
a plurality of subtiles and to assign at least some of those
subtiles to different processing cores for rendering. The
subdivision of tiles can be particularly useful for expensive tiles
occurring near the end of a render to reduce the impact on the
total render time when expensive tiles are scheduled near the end
of a render.
There is provided a graphics processing unit configured to process
graphics data using a rendering space which is sub-divided into a
plurality of tiles, the graphics processing unit comprising: a
plurality of processing cores configured to render graphics data;
cost indication logic configured to obtain a cost indication for
each of a plurality of sets of one or more tiles of the rendering
space, wherein the cost indication for a set of one or more tiles
is suggestive of a cost of processing the set of one or more tiles;
and scheduling logic configured to assign, in dependence upon the
cost indications, the sets of one or more tiles to the processing
cores for rendering.
Furthermore, there is provided a method of processing graphics data
in a graphics processing system which comprises a plurality of
processing cores configured to render graphics data, the graphics
processing system being configured to use a rendering space which
is sub-divided into a plurality of tiles, the method comprising:
obtaining a cost indication for each of a plurality of sets of one
or more tiles of the rendering space, wherein the cost indication
for a set of one or more tiles is suggestive of a cost of
processing the set of one or more tiles; and assigning, in
dependence upon the cost indications, the sets of one or more tiles
to the processing cores for rendering.
For example, the cost indication for a set of one or more tiles may
be based on one or more factors which influence a cost of
processing the set of one or more tiles. The scheduling logic may
assign the sets of one or more tiles for a render to the processing
cores in an order depending upon the cost indications.
The cost indication logic may be configured to determine the cost
indications for the sets of one or more tiles of the rendering
space. For example, the cost indication logic may be configured to
determine a cost indication for a tile of the rendering space by
combining scores associated with primitives which are present in
the tile, wherein the score associated with a primitive is
dependent upon an object type of an object of which the primitive
is a part. To give some more examples, the cost indication logic
may be configured to determine a cost indication for a tile of the
rendering space based on one or more of the following factors: (i)
a number of primitives in the tile; (ii) object types associated
with the primitives in the tile; (iii) tile coverage area of the
primitives in the tile; (iv) characteristics of one or more shader
programs which are to be executed for rendering the primitives in
the tile; (v) a user input; and (vi) a processing cost of a
corresponding tile in a previous render. The cost indication logic
may be configured to quantise the cost indications.
The sets of one or more tiles may each comprise a single tile. In
other examples, the sets of one or more tiles may each comprise a
plurality of tiles. The sets of tiles may be blocks of tiles of the
rendering space (i.e. the sets of tiles may be spatially
contiguous).
The graphics processing unit may comprise geometry processing logic
and rendering logic, wherein the geometry processing logic includes
a tiling unit configured to generate control streams for the tiles
of the rendering space indicating which primitives are present in
the tiles, wherein the cost indication logic is implemented as part
of the geometry processing logic, wherein the rendering logic
comprises the processing cores and is configured to render
primitives in tiles of the rendering space in accordance with the
generated control streams, and wherein the scheduling logic
determines the assignment of sets of one or more tiles to the
processing cores of the rendering logic. The tiling unit may be
configured to include the cost indications in the control streams
for the tiles.
There is provided a graphics processing unit configured to process
graphics data using a rendering space which is sub-divided into a
plurality of tiles, the graphics processing unit comprising: one or
more processing cores configured to process graphics data; cost
indication logic configured to obtain a cost indication for each of
a plurality of sets of one or more tiles of the rendering space,
wherein the cost indication for a set of one or more tiles is
suggestive of a cost of processing the set of one or more tiles;
and scheduling logic configured to schedule, in dependence upon the
cost indications, the sets of one or more tiles for processing on
the one or more processing cores.
Furthermore, there is provided a method of processing graphics data
in a graphics processing system which comprises one or more
processing cores configured to process graphics data, the graphics
processing system being configured to use a rendering space which
is sub-divided into a plurality of tiles, the method comprising:
obtaining a cost indication for each of a plurality of sets of one
or more tiles of the rendering space, wherein the cost indication
for a set of one or more tiles is suggestive of a cost of
processing the set of one or more tiles; and scheduling, in
dependence upon the cost indications, the sets of one or more tiles
for processing on the one or more processing cores.
Said one or more processing cores may comprise a plurality of
processing cores configured to render the graphics data.
Furthermore, the scheduling logic may be configured to assign, in
dependence upon the cost indications, the sets of one or more tiles
to the processing cores for rendering. The cost indication for a
set of one or more tiles may be based on the content of the set of
one or more tiles.
In some embodiments, there is provided a graphics processing unit
configured to process graphics data using a rendering space which
is divided into a plurality of tiles, the graphics processing unit
comprising: one or more processing cores configured to render
graphics data; and scheduling logic configured to schedule sets of
one or more tiles for rendering on the one or more processing
cores, wherein the scheduling logic is configured to, for at least
one of the sets of one or more tiles of the rendering space,
subdivide the set of one or more tiles to determine a plurality of
subunits, wherein the scheduling logic is configured to schedule
the subunits for rendering on the one or more processing cores. For
example, there may be provided a graphics processing unit
configured to process graphics data using a rendering space which
is divided into a plurality of tiles, the graphics processing unit
comprising: a plurality of processing cores configured to render
graphics data; and scheduling logic configured to assign sets of
one or more tiles to the processing cores for rendering, wherein
the scheduling logic is configured to, for at least one of the sets
of one or more tiles of the rendering space, subdivide the set of
one or more tiles to determine a plurality of subunits, wherein the
scheduling logic is configured to assign at least some of the
subunits derived from the same set of one or more tiles to
different processing cores for rendering.
The graphics processing units described herein may be embodied in
hardware on an integrated circuit. There may be provided a method
of manufacturing, at an integrated circuit manufacturing system, a
graphics processing unit as described herein. There may be provided
an integrated circuit definition dataset that, when processed in an
integrated circuit manufacturing system, configures the system to
manufacture a graphics processing unit as described herein. There
may be provided a non-transitory computer readable storage medium
having stored thereon a computer readable description of an
integrated circuit that, when processed, causes a layout processing
system to generate a circuit layout description used in an
integrated circuit manufacturing system to manufacture a graphics
processing unit as described herein.
There may be provided an integrated circuit manufacturing system
comprising: a non-transitory computer readable storage medium
having stored thereon a computer readable integrated circuit
description that describes a graphics processing unit as described
herein; a layout processing system configured to process the
integrated circuit description so as to generate a circuit layout
description of an integrated circuit embodying the graphics
processing unit; and an integrated circuit generation system
configured to manufacture the graphics processing unit according to
the circuit layout description.
There may be provided computer program code for performing any of
the methods described herein. There may be provided non-transitory
computer readable storage medium having stored thereon computer
readable instructions that, when executed at a computer system,
cause the computer system to perform any of the methods described
herein.
The above features may be combined as appropriate, as would be
apparent to a skilled person, and may be combined with any of the
aspects of the examples described herein.
BRIEF DESCRIPTION OF THE DRAWINGS
Examples will now be described in detail with reference to the
accompanying drawings in which:
FIG. 1 shows a prior art graphics processing system;
FIG. 2a is a graph illustrating an idealised tile workload
distribution on a 1 core GPU and a 10 core GPU;
FIG. 2b is a graph illustrating a tile workload distribution on a 1
core GPU and a 10 core GPU with small variations in the costs of
the tile workloads;
FIG. 3 shows a graphics processing system;
FIG. 4 is a flow chart for a method of processing graphics data
using the graphics processing system;
FIG. 5a shows sets of tiles of a rendering space;
FIG. 5b illustrates cost indications for the sets of tiles of the
rendering space;
FIG. 6 illustrates the timing of execution of the sets of tiles on
a 6 core GPU when a workload-based scheduling method is used which
preferentially fills empty cores first;
FIG. 7 illustrates the timing of execution of the sets of tiles on
a 6 core GPU when a scheduling method based on the cost indications
is used to assign the sets to the processing cores;
FIG. 8a illustrates a workload distribution over 32 cores for a
tile size of 32.times.32 sample positions;
FIG. 8b illustrates a workload distribution over 32 cores for a
tile size of 16.times.16 sample positions;
FIG. 9 is a flow chart for a method of processing graphics data
using the graphics processing system in a further example;
FIG. 10a illustrates a process of subdividing a tile in a first
example;
FIG. 10b illustrates a process of subdividing a tile in a second
example;
FIG. 11 shows a computer system in which a graphics processing
system is implemented; and
FIG. 12 shows an integrated circuit manufacturing system for
generating an integrated circuit embodying a graphics processing
system.
The accompanying drawings illustrate various examples. The skilled
person will appreciate that the illustrated element boundaries
(e.g., boxes, groups of boxes, or other shapes) in the drawings
represent one example of the boundaries. It may be that in some
examples, one element may be designed as multiple elements or that
multiple elements may be designed as one element. Common reference
numerals are used throughout the figures, where appropriate, to
indicate similar features.
DETAILED DESCRIPTION
The following description is presented by way of example to enable
a person skilled in the art to make and use the invention. The
present invention is not limited to the embodiments described
herein and various modifications to the disclosed embodiments will
be apparent to those skilled in the art.
Embodiments will now be described by way of example only.
In the graphics processing system shown in FIG. 1, increasing the
number of processing cores will tend to improve the performance of
the graphics processing unit in terms of reducing the time taken to
process all of the tiles of a render because the processing of the
tiles can be divided amongst the processing cores. However,
increasing the number of processing cores tends to increase the
size (i.e. silicon area) of the GPU and the peak processing power
consumed by the GPU.
FIG. 2a is a graph illustrating an idealised tile workload
distribution on a 1 core GPU and a 10 core GPU. The vertical axis
in FIG. 2a has time running downwards, from a render start time.
The block 202 represents the work performed by a single processing
core if the GPU 102 uses one processing core 114 for rendering all
of the tiles of the rendering space. In this example, the render
end time is shown at 204. In this idealised example, the blocks
206.sub.0 to 206.sub.9 represent the work performed by the
processing cores if the GPU 102 uses ten processing cores 114 for
rendering all of the tiles of the rendering space. In this example,
the render end time is shown at 208. In this idealised example,
each tile incurs the same processing cost, such that the render
time with one core is ten times as long as the render time with ten
cores.
As an example, consider a frame which consists of only a single
render. This render may be 640.times.480 pixels in size and
therefore there may be 300 tiles (arranged as a 20.times.15 block
of tiles) in the frame, where the tiles are each 32.times.32 pixels
in size. If all tiles take an equal amount of time `T` to process,
then a one cluster GPU can be expected to take a time of
approximately 300T to complete the render; whereas a ten cluster
GPU can be expected to take a time of approximately 30T to complete
the render. Therefore in this idealised case, a ten cluster GPU is
ten times quicker than a one cluster GPU at completing a render. It
is noted that the terms "cluster" and "core" may be used
interchangeably herein, both of which refer to a processing unit
(including processing components such as arithmetic logic units)
which is configured to perform processing operations on incoming
data. Furthermore, the terms "frame" and "image" may be used
interchangeably herein.
However, the perfectly balanced workload shown in FIG. 2a is rarely
what happens in real graphics workloads. Allowing for even a small
random variation in the work of each tile can significantly modify
the overall runtime of the frame on multiple cores even when the
total workload in the frame remains constant as shown in FIG. 2b.
FIG. 2b is a graph illustrating a tile workload distribution on a 1
core GPU and a 10 core GPU with small variations in the costs of
the tile workloads. In FIG. 2b a third of the tiles are `low cost`
tiles (shown as unhatched workloads), a third of the tiles are
`medium cost` tiles (shown with diagonal hatching) and a third of
the tiles are `high cost` (shown with cross hatching). The `cost`
of a tile may refer to the amount of time taken to process the tile
(as in the example shown in FIG. 2b). In other contexts, the cost
of a tile may refer to the amount of processing resources used, the
amount of data transferred to/from memory, or the processing power
used for processing the tile, etc.
In FIG. 2b, the block 212 represents the work performed by a single
processing core if the GPU 102 uses one processing core 114 for
rendering all of the tiles of the rendering space. In this example,
the render end time is shown at 214, which is the same as render
end time 204 shown in FIG. 2a. In this example, the blocks
216.sub.0 to 216.sub.9 represent the work performed by the
processing cores if the GPU 102 uses ten processing cores 114 for
rendering all of the tiles of the rendering space. In this example,
the render end time is shown at 218. The dashed line 208 shows the
render end time in the idealised example of FIG. 2a. A render is
complete when all of the tiles of the render have been processed.
Before a current render can begin execution, a previous render may
need to be completed, i.e. the current render waits for the
previous render to complete before commencing execution of tiles of
the current render. It can be seen in FIG. 2b that the variation in
the processing costs of the tiles results in a longer render time
when multiple cores are used, i.e. render end time 218 is after
idealised render end time 208. A render may be for generating pixel
values of a frame, such that completing the render involves storing
pixel values in a framebuffer to represent a frame, which can then
be used in any appropriate manner, e.g. displayed on a display,
stored or transmitted, etc. In other examples, a render may not
result in a finished frame, e.g. the render may be a sub-render
which can be used in a subsequent rendering process for generating
a frame. For example, a sub-rendering could be a `render to
texture` such that the result of the rendering is a texture which
can then be applied to a subsequent rendering process for
generating an image. As other examples, a sub-rendering may be a
shadow map or an environment map for subsequent use in rendering a
frame.
GPUs with real applications may typically run many tiles per render
(e.g. 2000 tiles) and may contain many renders (e.g. 20). This
gives a lot of scope for random variation beyond that shown in FIG.
2b. When tiles are individually assigned to particular processing
cores, there may be a random probability of consecutively executing
slow tiles (i.e. tiles with high processing costs) on a single
processing core. In systems which assign groups of more than one
spatially-adjacent tile to a processing core, there may be a worse
than random probability of consecutively executing slow tiles being
assigned to a single core. As described in more detail below, this
is because slow tiles typically correspond to complex regions of
the scene/objects, so since the tiles in the group are
spatially-adjacent then if one tile in the group is expensive then
it is likely that the other tiles in the group will also be
expensive. The slowest core represents a limiting factor for the
render and therefore the core that finishes last determines the
time it takes to execute the render and by extension both the
framerate of the application and the efficiency of the hardware.
Averaged out over many thousands of tiles it might be expected that
a random allocation scheme would average out to be reasonably well
balanced, however that is often not true. There are three
fundamental reasons for this:
1) By design, tile based GPUs often aim to process spatially local
tiles on the same cluster in order to maximise the efficiency of
caches. Spatially local tiles commonly share lots of data,
including things like textures, instructions, shaders, etc. which
means it may be beneficial from a cache coherency perspective to
keep multiple neighbouring tiles within the same cluster. This
tends to mean that the worst case scheduling of back to back
expensive tiles on a single core is not only possible but
positively reinforced. 2) The workload in graphics applications
such as games is typically not evenly distributed across the frame.
Some regions of the frame may contain scene objects or effects that
are particularly time consuming to compute (i.e. have a high
processing cost). Common examples are translucent objects,
punch-through objects (objects with transparent areas defined by
textures or shaders), effects that require multiple layers to be
blended together and some shader effects that are computationally
expensive but applied only to some objects/regions of the scene. It
is noted that punch through is a technique where an object may or
may not have regions that are transparent, and the presence or
location of these transparent regions is not known until runtime
when visibility is either calculated or fetched from a texture. A
common and very expensive example is foliage where leaves are drawn
as simple polygons (e.g. rectangles or triangles) represented by
one or more primitives, and the leaf shape is generated by making
some parts of the polygon invisible, e.g. by making some primitive
fragments invisible within the one or more primitives representing
the polygon. 3) An expensive tile that is processed towards the end
of a render may cause a significant increase in the time taken for
the render, if there are too few other tiles remaining to allow
other processing cores to remain busy for the duration of the
processing of the expensive tile. Allowing these other cores to
become idle is an inefficient use of the processing resources that
increases the run time of the whole render.
Due to the nature of realtime graphics a constant stream of renders
is performed and the random variations of run time can result in
the undesirable effect of creating a stuttering effect in the
output presented to an end user. Stuttering occurs when a screen
updates the display at a particular frequency (e.g. 60 Hz). If the
frame to frame variations mean that sometimes an updated frame is
available before the screen updates (resulting in smooth motion),
but at other times it is not available (resulting in a previous
frame being repeated) there will be a perceptible stuttering in the
final output images.
In examples described herein, work is scheduled for processing
cores of a multi-core GPU based on cost estimates for the different
items of work. In particular, for each tile of a render a cost
indication is determined, and then tiles can be assigned to the
processing cores based on the cost indications. A cost indication
for a tile indicates, or at least suggests, a cost of processing
the tile. For example, a cost indication may indicate a likely cost
of processing the tile. Scheduling the tiles in this manner can
reduce the likelihood of starting the processing of a high cost
tile near the end of a render. As described above, starting to
process high cost tiles near the end of a render can be
particularly detrimental to the overall render time, so reducing
the likelihood of this occurring can improve (i.e. reduce) the
average render time. Therefore, sets of tiles with relatively high
cost indications are preferentially scheduled before sets of tiles
with relatively low cost indications. In this description, a high
cost indication means a high processing cost, and a low cost
indication means a low processing cost. In some examples, the cost
indication for a set of one or more tiles is indicative of one or
more factors which influence a cost of processing the set of one or
more tiles.
FIG. 3 shows some elements of a graphics processing system 300
which may be used to render an image of a 3D scene. The graphics
processing system 300 comprises a graphics processing unit (GPU)
302 and two portions of memory 304.sub.1 and 304.sub.2. The two
portions of memory 304.sub.1 and 304.sub.2 may, or may not, be
parts of the same physical memory. The GPU 302 comprises a
pre-processing module 306, a tiling unit 308 and rendering logic
310, wherein the rendering logic 310 comprises a fetch unit 312,
processing logic 313 which includes one or more processing cores
(314.sub.1 to 314.sub.6) and scheduling logic 316 which comprises
one or more rendering queues 318. The rendering logic 310 is
configured to use the processing cores 314 of the processing logic
313 to implement hidden surface removal (HSR) and texturing and/or
shading on graphics data (e.g. primitive fragments) for tiles of
the rendering space. The tiling unit 308 comprises cost indication
logic 320. The graphics processing system 300 also comprises a
compiler 322 configured to compile programs (e.g. shader programs)
to be executed on the GPU 302. The compiler 322 may write compiled
shader programs to an intermediate memory, wherein at runtime the
GPU 302 retrieves the compiled shader programs from the
intermediate memory, but for simplicity, the intermediate memory is
not shown in FIG. 3. In the example shown in FIG. 3 the rendering
logic 310 comprises six processing cores 314.sub.1 to 314.sub.6,
but in other examples any suitable number of processing cores may
be included in the rendering logic 310, e.g. in a range from 1 to
256, or even higher. The number of processing cores in the
rendering logic 310 may be adapted to suit the intended use of the
graphics processing system (e.g. a graphics processing system to be
used in a small mobile device which has tight constraints on
processing resources and silicon size may include a small number of
processing cores (e.g. 6 processing cores), whereas a graphics
processing system to be used in a large device such as a PC or
server which has less tight constraints on processing resources and
silicon size may include a larger number of processing cores (e.g.
128 processing cores)).
The operation of the graphics processing system 300 is described
with reference to the flow chart shown in FIG. 4. Graphics data for
performing a render is received at the GPU 302, e.g. from a game
application running on a CPU in the same computing system as the
GPU 302. The graphics data may include primitive data describing
primitives of objects in a scene to be rendered. The graphics data
may also specify one or more shader programs which are to be
executed on the primitive data for rendering the primitives. It is
noted that shaders specifying position (e.g. vertex and geometry
shaders) are executed on primitives; whereas shaders specifying how
the image will be rendered (e.g. pixel shaders) are executed on
primitive fragments corresponding to parts of primitives that cover
pixels (or more precisely that cover sample positions).
In step S402 the compiler 322 compiles the shader programs which
are associated with the primitives and determines characteristics
of the shader programs. The characteristics which are determined
include characteristics which are indicative of the complexity of
the shader program. In particular, the characteristics are
determined so as to give an indication of a cost of processing
primitives using the particular shader programs. For example, the
length of the shader program (e.g. number of operations to be
performed) may be identified. Furthermore, shader programs are
identified as being potentially high cost if they contain loops
that execute for a variable number of times, where that variable
number is determined at runtime, i.e. it is not a known number at
compilation time. Loops such as this are potentially very costly to
execute if they loop a large number of times in runtime. As another
example, the compiler could determine an amount of resources,
memory reads or registers used by the shader program, and use this
as a measure of the processing cost associated with running the
shader program. As an example, a shader which involves lots of
sampling from textures which may be sampled from external memory
will likely take significantly longer to process than one that
primarily consists of arithmetic instructions. Other
characteristics which may be determined include whether the shader
program includes conditional flow control.
Step S402 may be performed prior to runtime, i.e. in an offline
process. For example, the shader programs may be compiled when the
application loads. In particular, the shader programs may be
compiled before the rendering begins (e.g. before the tiling phase
begins) and before the shaders are associated with any specific
geometry. However, in other examples it would be possible for a
shader program to be compiled during runtime before the main
rendering begins, e.g. in response to determining that a primitive
is to be processed using the shader program. The compiler 322 can
flag a wide number of potentially expensive things that may be
present in a shader program. The compiler 322 is arranged to
provide the determined characteristics of the shader programs to
the tiling unit 308. The compiler 322 is arranged to provide the
compiled shader programs to the rendering logic to be executed on
one or more of the processing cores 314 for processing
primitives.
A sequence of primitives provided by an application may be received
at the pre-processing module 306. In a geometry processing phase,
the pre-processing module 306 performs functions such as geometry
processing including clipping and culling to remove primitives
which do not fall into a visible view. The pre-processing module
306 may also project the primitives into screen-space. The
primitives which are output from the pre-processing module 306 are
passed to the tiling unit 308 for tiling as described below.
In step S404 the tiling unit 308 determines which primitives are
present within each of the tiles of the rendering space of the
graphics processing system 300. The tiling unit 308 assigns
primitives to tiles of the rendering space by creating control
streams for the tiles, wherein the control stream for a tile
includes indications of primitives which are present within the
tile. The control streams and the primitives are outputted from the
tiling unit 308 and stored in the memory 304.sub.1. The geometry
processing phase (performed by the pre-processing module 306 and
the tiling unit 308) takes account of primitives across the whole
of an image, i.e. for all of the tiles in the image. Then in the
rendering phase, the rendering logic 310 renders tiles of the image
and stores the outputs for rendered tiles in appropriate portions
of a framebuffer, such that when all of the tiles of an image have
been rendered, the framebuffer stores the rendered results for the
whole image. In examples described herein, the opportunity that is
provided in the geometry processing phase to assess all of the data
for an image before tiles are rendered for the image is used to
determine information about the image which may be useful for the
rendering phase, e.g. to improve the efficiency of the rendering
phase. In examples described below, tiles can be scheduled for
processing by the rendering logic based on an estimate of the
processing cost that will be involved in processing the tiles. This
can reduce the likelihood of scheduling high cost tiles near the
end of a render.
In step S406 the cost indication logic 320 determines cost
indications for the tiles of the rendering space. As described
above, the cost indication for a tile suggests a cost of processing
the tile. The cost indication logic may determine the cost
indications based, at least in part, on the determined
characteristics of the shader programs that were determined by the
compiler 322 in step S402. Furthermore, in general as described
below, cost indications may be determined for sets of one or more
tiles, i.e. a cost indication may be determined for a tile and/or a
cost indication may be determined for a set of tiles. For
simplicity some of the explanation herein refers to there being a
cost indication for a tile, but in general it is to be understood
that this explanation could be extended to having a cost indication
for a set of tiles.
The cost indications may be different in different examples. In
some examples, the cost indication for a tile may be an estimate of
a processing cost that will be incurred when the tile is processed
by the rendering logic 310. As described above, a processing cost
could be a length of processing time, a number of computation
operations performed, a processing power consumed, a number of
reads/writes from/to memory, or any other suitable measure of the
cost of processing a tile. However, in some examples, the cost
indication for a tile might not be a direct estimate of a
processing cost. The cost indication for a set of one or more tiles
may be based on the content of the set of one or more tiles. The
cost indication for a set of one or more tiles may be based on one
or more factors which influence a cost of processing the set of one
or more tiles. For example, a cost indication could be a number of
primitives which are present in a tile. The number of primitives in
a tile is not a direct estimate of the cost of processing the tile,
but it is indicative of an approximate processing cost that is
likely to be involved in processing a tile. For example, a larger
number of primitives in a tile may suggest that the tile will incur
a greater processing cost. In a broad sense, the cost indication
for a tile could be any parameter which is suggestive of a cost of
processing the tile, i.e. a parameter which provides some measure
of likely processing cost for use in distinguishing between tiles.
It is further noted that the cost indications might not always
accurately reflect the true processing costs of processing tiles,
but they aim to provide a better indication of processing costs for
tiles than if no cost indications were determined at all.
In a simple example, the cost indication for a tile is the number
of primitives which are present in the tile. A tile which overlaps
with a relatively large number of primitives tends to incur a
greater processing cost than a tile with a relatively small number
of primitives, so the number of primitives in a tile is a useful
cost indication even if it does not always reflect the exact actual
processing cost of rendering the tile. Furthermore, the number of
primitives in a tile is very simple to calculate in the tiling unit
308 because it can be directly observed from the control stream for
the tile, i.e. the number of primitive identifiers included in the
control stream for a tile at the end of the tiling phase indicates
the number of primitives in that tile. So in this example, the cost
indication logic 320 does not add significant complexity to the
tiling unit 308.
In a slightly more complex example, the cost indication logic 320
determines the cost indication for a tile by combining scores
associated with primitives which are present in tile. The score
associated with a primitive may be dependent upon an object type of
an object of which the primitives is a part. For example,
primitives associated with an opaque object type may be relatively
simple to process in the rendering logic 310, so these primitives
may be associated with low scores; whereas primitives associated
with other object types, e.g. translucent or punch through object
types or object types allowing primitives to change their depths
during rendering, may be relatively complex to process in the
rendering logic 310, so these primitives may be associated with
high scores. In particular, the rendering of these more complex
object types (e.g. translucency and punch through and types
allowing objects to change depth during rendering) may utilise
blending or other operations that require multiple passes in the
rendering logic 310 to resolve the pixels covered by these
primitives. For example, each primitive associated with an opaque
object type may be given a score of one, each primitive associated
with a translucent or punch through object type may be given a
score of ten, and each primitive which may change depth during
rendering may be given a score of eight. This reflects a likely
difference in the processing costs of the different types of
primitives. The scores for the primitives within a tile can be
summed, or combined in another way, to provide a cost indication
for the tile. In different examples, the scores for different
object types may be different to those described herein.
Tessellation is a technique which allows a graphics data item
(which may be referred to as a "patch") to be expanded into many
primitives during rendering. Tessellation can be useful for
representing complex (e.g. curved) surfaces, but can result in a
large number of primitives being rendered. A cost indication for a
tile could be based on whether tessellation is applied to patches
in the tile. As an example, if a tile includes a patch to be
tessellated, a cost indication for the tile could depend upon the
number of triangles which result from the tessellation of the
patch.
In other examples the cost indications may be determined in
different ways. For example, the tile coverage area of the
primitives in a tile may be considered when determining the cost
indication for the tile. The tile coverage area of a primitive
indicates a number of sample positions at which that primitive may
be visible within the tile, and therefore provides an indication of
the amount of processing that will be performed when processing the
primitive in the tile in the rendering logic 310. In some examples,
a user could provide a user input to guide the determination of the
cost indications for the tiles. In this sense the cost indication
logic 320 may receive the cost indications for the tiles via an
input. For example, a user may be able to specify the cost
indications for the tiles directly, e.g. via an API extension, to
allow a developer to explicitly provide tile costs to ensure
efficient performance using a priori knowledge of the workloads
associated with particular tiles.
A driver mechanism may pass information from the compiler 322 to
the tiling unit 308, and this information may include the
characteristics of the shader programs determined by the compiler
322. Optionally the driver may wish to be used to flag geometry
that must be regenerated (e.g. pipeline stages such as geometry
shaders and tessellation shaders can be used to expand primitives
to create multiple primitives), and in some graphics processing
systems the expanded primitives are not stored after the geometry
processing phase and must be regenerated before use in the
rendering phase. Similarly the driver may also provide information
on the frequency at which 3D shading will be performed, e.g. it is
possible for the rendering phase to render at a higher pixel rate
or a higher sample/fragment rate which is a more costly process.
Therefore this information can be useful for the cost indication
logic 320 for determining the cost indications.
The examples described above relate to the factors relating to the
processing of the current render which can be used to estimate
likely processing costs for rendering different tiles of the
render. As well as these factors, the cost indication logic 320
could determine the processing costs (either predicted or actual
costs) for tiles of a previous render (e.g. the immediately
preceding render, e.g. the preceding frame), and can use these as a
factor in determining the cost indications for the tiles of the
current render. Two frames of a sequence of frames are likely to be
similar if they are close to each other in the sequence, e.g. if
they are consecutive frames, unless there is a scene change or a
sudden change in the content. Therefore, the processing costs of
particular tiles in a previous frame provide a good indication of
the processing costs of corresponding tiles in a current frame. The
"corresponding tiles" in different frames may be tiles in the same
position within the rendering space, or may be displaced relative
to each other, e.g. by an amount representative of motion of
content in the scene (e.g. represented by motion vectors).
To summarise some of the examples described above, the cost
indication logic 320 may determine a cost indication for a tile of
the rendering space based on one or more of the following factors:
(i) a number of primitives in the tile; (ii) object types
associated with the primitives in the tile; (iii) tile coverage
area of the primitives in the tile; (iv) characteristics of one or
more shader programs which are to be executed for rendering the
primitives in the tile; (v) a user input; and (vi) a processing
cost of a corresponding tile in a previous render. However, it will
be apparent that other factors may be used in other examples for
determining the cost indications. The cost indication logic 320 may
determine the cost indication for a tile based on a plurality of
the factors, e.g. according to any suitable combination, which may
or may not be weighted in favour of one factor over another
factor.
As described above, the characteristics of a shader program may
include one or more of: (i) a length of the shader program; (ii) an
amount of resources or registers used by the shader program; (iii)
whether the shader program includes conditional flow control; (iv)
whether the shader program includes loops for which the number of
repetitions is undefined at compile time; and (v) a number of
memory reads and/or writes used in the shader program.
As described above a cost indication may be determined for each set
of one or more tiles. It may be the case that each set of one or
more tiles comprises the same number of tiles. In some examples the
sets of one or more tiles each comprise a single tile. In other
examples, the sets of one or more tiles each comprise a plurality
of tiles. The sets of tiles may be blocks of tiles (e.g. contiguous
tiles) of the rendering space. The term "block" of tiles is used
herein to refer to a plurality of spatially local or adjacent
tiles. In particular, the sets of one or more tiles may be arranged
to match the assignment of sets of tiles to processing cores 314 in
the rendering logic 310. For example, if individual tiles are
assigned to particular ones of the processing cores 314 at a time
then the sets of tiles may comprise single tiles. However, if
blocks of multiple tiles (e.g. 2.times.2, 4.times.2 or 4.times.4
blocks of tiles) are assigned to particular ones of the processing
cores 314 at a time then the sets of tiles may comprise
corresponding blocks of tiles. As described below, it may be
efficient from a cache coherency perspective to assign blocks of
tiles to processing cores 314 rather than assigning individual
tiles to processing cores 314.
The cost indication logic 320 may quantise the cost indications. In
particular, the quantised cost indications may be quantised to be
represented by a number of bits, the number of bits being in a
range from 1 to 8. In an extreme example, the quantised cost
indications each have a single bit, such that they act as a flag to
indicate that a tile is either a high cost tile or a low cost tile.
Even when the cost indications are quantised to this extent the use
of the cost indications can be useful for scheduling the processing
of the tiles because it will tend to avoid situations where a high
cost tile is scheduled for processing near the end of a render,
which as described above can cause a particularly long delay in the
render time. Quantising the cost indications reduces the amount of
data used to store the cost indications, and as explained below in
some examples simplifies the scheduling logic 316 by reducing the
number of priority queues implemented therein.
In step S408 the tiling unit 308 includes the determined cost
indications in the control streams for the tiles to be stored in
the memory 304.sub.1. For example the tiling unit 308 could include
the determined cost indications in a tail pointer cache which
includes pointers to the ends of the lists of primitive IDs for
respective tiles. The pointers can be included in the tail pointer
cache for the tiles of a rendering space at the end of processing
the geometry for a render in the tiling unit 308, and at this point
the cost indication logic 320 has determined the cost indications
for the tiles, so this is a good opportunity to add the cost
indications in a suitable place in the control streams without
having to significantly alter the operation of the tiling unit 308.
At the end of the geometry processing phase the control streams for
the tiles (including the cost indications) are stored in the memory
304.sub.1.
In the rendering phase the scheduling logic 316 receives the
control streams for the tiles of the rendering space for a current
render. The scheduling logic 316 is shown as being part of the
rendering logic 310, but in other examples the scheduling logic 316
may be thought of as a tile sequencer which acts outside of the
rendering logic 310 and which determines the order of the tiles to
be processed by the rendering logic 310. In some examples, the
scheduling logic could be implemented as part of the fetch unit
312.
In step S410 the scheduling logic 316 performs a pre-pass of the
control streams in order to analyse the cost indications for the
tiles to assign the tiles to the processing cores 314. In this way
the scheduling logic assigns, in dependence upon the cost
indications, the sets of one or more tiles to the processing cores
for rendering. For example, the scheduling logic 316 may assign
sets of one or more tiles for the render to the processing cores
314 in an order depending upon the cost indications. In another
example, the scheduling logic 316 may assign a plurality of sets of
one or more tiles with the highest cost indications to different
processing cores 314, to thereby distribute the sets of tiles with
the highest cost indications amongst the processing cores 314.
In examples described herein the scheduling logic 316 schedules the
sets of tiles for processing on the processing cores 314 in an
order that executes relatively high cost tiles before relatively
low cost tiles. It is possible to do this because, in a tile based
system, each tile is rendered independently of all other tiles. For
example, the scheduling logic 316 may maintain a rendering queue
318 for each of the processing cores 314 to indicate which tiles
are to be processed by the respective processing cores 314, and the
order in which they are to be processed.
The scheduling logic 316 may form a priority queue (different to
the rendering queues 318) for sets of tiles for each of the
different cost indication values. This works particularly well when
the cost indications have been quantised such that there are a
small number of distinct cost indication values. The sets of one or
more tiles are assigned to the processing cores for rendering in an
order according to the contents of the priority queues. For
example, if the cost indications are quantised down to 2-bit
values, there are four different quantised cost indication values
(0, 1, 2 and 3) and hence the scheduling logic 316 maintains four
priority queues. Indications of sets of tiles which have a
quantised cost indication value of zero are stored in priority
queue 0; indications of sets of tiles which have a quantised cost
indication value of one are stored in priority queue 1; indications
of sets of tiles which have a quantised cost indication value of
two are stored in priority queue 2; and indications of sets of
tiles which have a quantised cost indication value of three are
stored in priority queue 3. When the scheduling logic 316 comes to
schedule a set of tiles for processing (e.g. by adding a set to the
rendering queue 318 associated with one of the processing cores
314) the scheduling logic 316 will select a set which is identified
in priority queue 3 unless priority queue 3 is empty, in which case
the scheduling logic 316 will select a set which is identified in
priority queue 2 unless priority queue 2 is empty, in which case
the scheduling logic 316 will select a set which is identified in
priority queue 1 unless priority queue 1 is empty, in which case
the scheduling logic 316 will select a set which is identified in
priority queue 0 unless priority queue 0 is empty, in which case
there are no more sets to schedule.
In step S412 the rendering logic 310 renders the tiles using the
processing cores 314 according to the determined scheduling. Steps
S410 and S412 may be performed concurrently such that sets of tiles
are assigned to processing cores as the processing cores render
other, previously assigned, sets of tiles. As described above, in
this example, the processing of a particular tile is performed by a
single one of the processing cores 314, i.e. the processing of a
tile is not divided between multiple processing cores 314. This
helps to improve the cache coherency and efficiency of processing
by reducing a need to communicate data between the processing cores
314. Step S412 involves the fetch unit 312 fetching the primitive
data for rendering a tile and providing the data to the appropriate
one of the processing cores 314 (based on the scheduling determined
by the scheduling logic 316). The processing logic 313 uses the
processing cores 314 to execute operations on the primitive data
for a tile in order to render the tile. The processing logic 313
includes other components (not shown in FIG. 3) such as registers,
caches, control logic, etc. for performing the processing of the
primitive data using the processing cores 314. Caches may be used
in the processing logic 313 to store data used by the processing
cores 314. Any suitable cache structure could be used (e.g.
multi-level or single level), and in particular, there may be some
caches which are coupled to at least one, but not all, of the
processing cores 314. To give some examples, each processing core
314 could have its own dedicated cache, or each pair of processing
cores could have their own dedicated cache.
The processing logic 313 may perform deferred rendering such that
hidden surface removal is performed on primitives to remove
primitive fragments which are hidden from view by other primitives
in the scene, and then texturing and/or shading is applied to the
primitive fragments after the hidden surface removal has been
applied to those primitive fragments. Texturing typically involves
reading texture data from a memory and applying the texture to
primitive fragments in order to determine colour values of rendered
primitive fragments. Shading typically involves executing shader
programs on primitive fragments to add visual effects to the
rendering of the primitive fragment. These shader programs are the
shader programs mentioned above that were compiled by the compiler
322 and provided to the rendering logic 310. Methods of performing
hidden surface removal, texturing and shading are known in the art
and as such the details of these methods are not described in
detail herein.
In other examples, the processing logic 313 may perform
non-deferred rendering such that texturing and/or shading is
applied to primitive fragments, and then hidden surface removal is
performed on the textured/shaded primitive fragments to remove
primitive fragments which are hidden from view by other primitives
in the scene. Non-deferred rendering methods may be less efficient
than deferred rendering methods because they involve unnecessarily
shading and texturing of primitive fragments which are ultimately
hidden in the scene.
Both the deferred rendering and non-deferred rendering systems
described above implement rasterisation techniques to render
primitive data. In other examples, other rendering techniques may
be used in the rendering logic, for example a ray tracing technique
may be used to render the primitive data. Ray tracing techniques
are known in the art and as such the details of implementing a ray
tracing rendering technique are not described in detail herein.
In step S414 the results of the render are stored. For example, if
the result of the render is a frame then the rendered frame may be
stored in the memory 304.sub.2. A rendered frame may be used in any
suitable manner. For example, a rendered frame may be displayed on
a display. A rendered frame may be transmitted to another device,
e.g. over a network such as the Internet and/or a mobile telephone
network. If the render is a sub-render then the result of the
render may be for use in a subsequent render, e.g. the result of
the sub-render could be a texture, shadow map or environment to be
applied in a subsequent render. In this case, the result of the
render could be stored (e.g. in memory 304.sub.2) and subsequently
provided back to the rendering logic 310 for use in the subsequent
render. Alternatively, the result of the render could be stored on
the GPU 302 itself.
An example is now described with reference to FIGS. 5a to 7. FIG.
5a shows a rendering space 502 which has 896.times.512 pixels. Each
tile is a 32.times.32 block of pixels, such that the rendering
space 502 has 28 columns and 16 rows of tiles as shown in FIG. 5a.
The tiles are grouped together into 4.times.4 blocks as shown by
the bold lines in FIG. 5a, wherein each 4.times.4 block of tiles is
to be processed by a particular processing core 314. By processing
4.times.4 blocks of tiles on the same processing core, the cache
hit rate of caches in the processing logic 313 which are used by
that processing core 314 may be improved since primitive data that
is relevant for one tile is more likely to be relevant for a nearby
(e.g. adjacent) tile than it is to be relevant for a distant tile
of the rendering space.
In this example, the cost indication logic 320 obtains cost
indications for each of the tiles in the rendering space, either by
determining the cost indications or by receiving the cost
indications via an API as described above. A cost indication is
then determined for each of the 4.times.4 sets of tiles based on
the cost indications of the tiles within the respective sets. To
give some examples, the cost indication logic may determine the
cost indication for a set of tiles by determining: (i) an average
of the cost indications of the tiles within that set, (ii) a sum of
the cost indications of the tiles within that set, (iii) a maximum
of the cost indications of the tiles within that set, or (iv) a
number of the tiles within that set which have cost indications
above a threshold. In other examples, the cost indications of tiles
within a set may be combined in other ways to determine a cost
indication for the set as a whole. Furthermore, in some examples, a
cost indication is determined for a set of tiles without first
determining individual cost estimates for the individual tiles
within the set.
FIG. 5b shows cost indications for the 4.times.4 sets of tiles
shown in FIG. 5a. In this example, cost indications are determined
for each of the tiles and then the cost indication for a 4.times.4
set of tiles is determined by combining the cost indications for
the individual tiles within that set of tiles (e.g. by finding the
sum of the individual cost indications) and then quantising the
cost indications down to a 4-bit value. 4-bit values are in a range
from 0 to 15. It can be seen that there is a region near the bottom
right of the rendering space 502 which is relatively costly to
render, e.g. because the graphics content in this region is
particularly detailed (e.g. a high number of primitives in this
region, which may involve translucent or punch through primitives)
or the effects used to render this region require complex shader
programs. To give some examples, regions of foliage, smoke, fire or
other particle systems in a scene can be costly to render. In
contrast the top left region of the rendering space 502 is not so
costly to render, e.g. because the graphics data in this region is
simple, e.g. this region may represent a plain background with
little detail in the scene.
Each set of tiles is to be fully processed by one processing core
and in this example the cost indications for the sets of tiles
indicate the relative time taken to complete the processing of each
set of tiles. If the sets of tiles are to be assigned to the
processing cores, a simple scheduler may allocate the sets of tiles
in a raster scan order according to a round robin scheme across
each of the cores in turn.
An improvement upon this round robin scheduling method is to
preferentially schedule work to processing cores that have the
least work queued up. This scheduling method is referred to herein
as a "workload-based scheduling method". For example, the
scheduling logic 316 may add sets of tiles to respective rendering
queues for the processing cores 314 during rendering in accordance
with the occupancy of the queues. FIG. 6 illustrates the timing of
execution of the sets of tiles on a 6 core GPU when a
workload-based scheduling method is used which preferentially
schedules work to processing cores that have the least work queued
up. In the simplified example shown in FIG. 6 the quantized costs
for the sets of tiles shown in FIG. 5b are used to represent the
actual processing times. The example shown in FIG. 6 does not use
the cost indications to schedule the rendering of the tiles. The
example shown in FIG. 6 is a simplified example to aid the
explanation of the scheduling within the system in which sets of
tiles are allocated to particular processing cores when the
processing cores become idle. In this simplified example, the
rendering queues are not necessary. However, in more complex
systems, the rendering queues are used to store work which has been
allocated to a core before the core is ready to process it. For
example, when a processing core is processing a set of tiles but
has no sets of tiles queued in its respective rendering queue, then
a set of tiles may be scheduled for the core and added to the
respective rendering queue. In this way the scheduling operates
slightly less optimally than if the sets of tiles were scheduled
for execution on a processing core when that processing core
becomes idle. However, by queuing at least one set of tiles for
each core in the rendering queues then some processing efficiencies
can be achieved, e.g. before a first set of tiles has completely
finished being processed by a processing core, data can be fetched
and prepared for a second set of tiles, such that the processing
core can immediately start processing the second set of tiles when
it has completed the processing of the first set of tiles.
The sets of tiles are allocated to the processing cores in a raster
scan order, so initially the first six sets of tiles from the top
row of the rendering space (tiles 0,0 to 5,0) are allocated to the
six processing cores 314.sub.1 to 314.sub.6. The sets of tiles take
an amount of time to process as shown in FIG. 6 matching the
respective cost indications for the sets. When a processing core
finishes processing a set of tiles the next set of tiles according
to the raster scan is allocated to that processing core (e.g. set
6,0 is allocated to processing core 314.sub.1 after time 2 when the
first set of tiles 0,0 have been processed). The next row of sets
of tiles (sets of tiles 0,1 to 6,1) are allocated according to the
raster scan when processing cores complete the processing of a set
of tiles. This scheduling continues as shown in FIG. 6 until after
time 31 all of the sets of tiles in the rendering space 502 for the
current render have been processed. Therefore, according to this
scheduling scheme the render is complete after 31 time intervals.
The time intervals could represent any suitable quantity of time or
indication of a quantity of time (e.g. thousands of clock cycles).
The hatched areas in FIG. 6 indicate times for which a processing
core is idle. In this example, there are a total of 31 idle time
slots across the processing cores for the render. It is noted that
in the examples described herein a subsequent render does not begin
to be processed by the processing cores before a current render has
finished. So, these 31 idle time slots represent inefficiency in
the graphics processing system, which may add delay to the renders
performed by the graphics processing system.
The workload-based scheduling scheme as shown in FIG. 6 is better
than a simple round robin scheduling scheme (in terms of rendering
efficiency) but it still has some inefficiency when the processing
cores are idle (shown with the hatched regions) caused by differing
amounts of work involved in the processing of different tiles. FIG.
6 illustrates that the amount of inefficiency (i.e. the number of
idle time slots) is largely determined by how late in the render
expensive/slow sets of tiles are received and the amount of
variation in the processing costs of tiles. These parameters are
not known in advance of runtime and depend upon the images
submitted for rendering by the application at run time.
FIG. 7 illustrates the timing of execution of the sets of tiles on
a 6 core GPU when a scheduling method based on the cost indications
is used to assign the sets to the processing cores. Using knowledge
about which sets of tiles are likely to be expensive can be useful
for scheduling the most expensive tiles first and use the
relatively cheap/fast tile groups to fill in the gaps later in the
render. The scheduling does not follow a raster scan; instead when
a set of tiles is due to be assigned to a processing core (e.g.
because the processing core finishes processing a set of tiles)
then a set of tiles is selected from the sets of tiles of the
render which have not yet been assigned to a processing core, the
selection being based on the cost indications for the sets of
tiles, rather than the spatial position of the sets of tiles in the
rendering space. For example, the set of tiles with the highest
cost indication of the remaining sets of tiles is assigned to a
processing core. This results in an overall more efficient usage of
the processing cores 314 as shown in FIG. 7, which shows the
processing of the same tiles as in FIG. 6, but this time the render
finishes after 27 time intervals, with fewer idle time slots. FIG.
7 shows just 7 idle time slots compared to the 31 idle times slots
in FIG. 6. Therefore the render finishes sooner with the scheduling
based on cost indications than with the scheduling shown in FIG. 6.
Therefore the scheduling based on cost indications can provide a
more efficient rendering process. A 13% reduction in the rendering
time is achieved by the scheduling of FIG. 7 compared to the
scheduling of FIG. 6 in the example shown (31-27/31=13%). The
reduction in rendering time is dependent upon the workload
(presence and distribution of expensive tiles) and the number of
processing cores in the system. Example 3D game and benchmark
content have shown reductions in overall frame render times in a
range from 0 to 30% (compared to the workload-based scheduling
shown in FIG. 6), with some outliers having even higher reductions.
This can be a very significant improvement particularly if the
graphics processing system 300 is rendering frames in real-time for
display in real-time to a user, and particularly if the graphics
processing system 300 is implemented on a device where the
processing resources are under a tight constraint, e.g. on a mobile
device such as a smart phone or tablet. The example shown in FIG. 7
is simplified in a similar manner to the example shown in FIG. 6
wherein sets of tiles are scheduled when processing cores become
idle, such that rendering queues are not used, but it is to be
appreciated that in more complex examples the rendering queues can
be implemented with a scheduling method which assigns the sets of
tiles to the processing cores based on the cost indications.
As mentioned above, even if the cost indications for tiles do not
perfectly indicate the actual cost of processing the tiles such
that a relatively cheap tile was scheduled early then the system
would still remain more balanced than a conventional approach as
the system would re-balance around it. So the cost indications do
not have to be estimated perfectly, to generate a good result in
terms of reducing the rendering time for a render. It is better to
catch all of the expensive tiles and accidently include tiles that
turn out to be cheaper than expected than to be conservative with
identifying expensive tiles.
In the example described with reference to FIGS. 5a to 7, the
scheduling logic 316 assigns the sets of tiles to the processing
cores 314 based on the cost indications for the sets of tiles.
Furthermore, the scheduling logic 316 may determine a processing
order of tiles within a particular set of tiles. For example, the
tiles within a particular set of tiles may be processed in raster
scan order or boustrophedon order, or according to Morton order
(which may be referred to as Z-order) or in accordance with a space
filing curve which preserves spatial locality between consecutively
processed tiles. Suitable space filling curves may include Peano
curves and Hilbert curves to give two examples. In some examples,
the tiles within a particular set of tiles may be processed in an
order based on the cost indications of the tiles within that
particular set.
In the examples described above, the cost indications for the sets
of one or more tiles are used to schedule the sets of one or more
tiles for rendering by the processing cores 314. In general, the
cost indications for the sets of one or more tiles could be used to
schedule the processing of the sets of one or more tiles on one or
more processing cores. In some examples, when selecting the next
tile or set of tiles to process, the scheduling logic 316 may use a
combination of cost indication and distance from a previously
scheduled tile (e.g. the last tile processed or scheduled for a
processing core 314), to find a balance between cache coherency and
scheduling optimisation. The relative importance of these factors
may vary as the render progresses.
The processing performed by the one or more processing cores could
be rendering as in the examples described above, or could be other
processing such as video encoding. In the examples described above,
the scheduling attempts to avoid starting the processing of an
expensive tile near the end of a render. In the examples described
above, the scheduling decision for a particular tile does not alter
the manner in which that particular tile is subsequently
rendered.
However, in other examples, the cost indications may be used for
other purposes. For example, when a frame has been rendered it may
be passed to a video encoder which encodes the frame according to a
video encoding technique. The encoded video data is generally a
compressed representation of the frames, which is suitable for
transmission (e.g. to a display for display thereon) and/or for
storage. As part of the video encoding, quantisation levels may be
selected for encoding regions of a frame. The cost indications for
the different tiles which are determined in the examples described
above may be used to help the video encoder select suitable
quantisation levels for encoding the different tiles. For example,
the cost indications for the sets of one or more tiles may be
provided to the video encoder with the rendered image data.
In other examples, the cost indications could be used to alter the
manner in which different tiles are processed, e.g. the manner in
which the tiles are rendered. For example, a high cost indication
may indicate that there is a lot of detail in a tile (e.g. when the
cost indications represent the number of primitives in a tile), in
which case these tiles could be rendered with a greater level of
detail (e.g. more sample positions per pixel could be processed in
a rasterisation technique or more rays could be processed per pixel
in a ray tracing technique). Alternatively, if there is a real-time
constraint on the timing with which tiles are processed then tiles
with a lot of detail (as indicated by the cost indications) may be
processed with a lower level of detail to ensure that they are
processed in time for the real-time constraint to be satisfied. For
example, images may be rendered and directly displayed according to
a raster scanning technique of the display. In this example, each
tile may have a deadline by which it must be rendered in order to
be displayed in real time as the raster scan of the display
proceeds. If a tile is not rendered before its deadline then it
might not be able to be displayed in a current frame, which can
cause a stuttering effect in the displayed images. The cost
indications provide some indication as to how long a tile will take
to process and as such can be used to vary the level of detail of
rendering a tile with the aim of ensuring that each tile is
rendered before its deadline for providing the rendered tile to a
display.
In general, a graphics processing unit which is configured to
process graphics data using a rendering space which is sub-divided
into a plurality of tiles may include one or more processing cores
configured to process graphics data; cost indication logic
configured to obtain a cost indication for each of a plurality of
sets of one or more tiles of the rendering space, wherein the cost
indication for a set of one or more tiles is suggestive of a cost
of processing the set of one or more tiles; and scheduling logic
configured to schedule, in dependence upon the cost indications,
the sets of one or more tiles for processing on the one or more
processing cores.
In some examples, which might not schedule tiles based on cost
indications for the tiles, the scheduling logic 316 may subdivide
one or more tiles for a render to form smaller subtiles for
rendering. As described in detail below, in some situations this
may help to reduce a total rendering time for a render. For
example, subdividing expensive tiles into subtiles can allow a
fairer distribution of work across multiple processing cores. The
application of this method is particularly beneficial towards the
end of each render and is an improvement upon simply reducing the
size of tiles throughout all renders. The cost of each tile is
directly determined by the size of the tile, i.e. smaller tiles
have less overall work in them and therefore complete more quickly.
The problem of long running tasks being scheduled towards the end
of a render is reduced and the smaller size of work packets
inherently allows for better load balancing using a
simple/conventional scheduler.
FIGS. 8a and 8b provide an illustration that having a smaller tile
size can reduce the rendering time for a render. FIG. 8a shows the
work done by 32 processing cores (denoted 802) in a graphics
processing system for rendering a frame using tiles which are
32.times.32 pixels in size. For each processing core the black line
shows when the processing core is active, as a function of time
increasing vertically downwards in the illustration. In the example
shown in FIGS. 8a and 8b tiles have not been scheduled for
rendering according to their cost indications, and as such a tile
having a high processing cost may be assigned to a processing core
near the end of the render. Instead in the example shown in FIGS.
8a and 8b the tiles are allocated according to a round robin
scheduling method which preferentially schedules work to processing
cores that have the least work queued up (similar to the example
shown in FIG. 6 and described above). Since the tiles are
relatively large in the example shown in FIG. 8a the work involved
in processing a tile may be relatively large if it involves lots of
costly processing, such as a large number of translucent primitives
or the execution of long shader program for example. Therefore it
can be seen in FIG. 8a that some of the processing cores are busy
processing expensive tiles at the end of the render whilst other
processing cores are idle because they have finished processing the
tiles that were assigned to them and there are no more tiles that
can be assigned to them. The render is complete at time 804.
FIG. 8b shows the work done by 32 processing cores (denoted 806) in
a graphics processing system when rendering the same frame as for
FIG. 8a, but this time using tiles which are 16.times.16 pixels in
size (i.e. a quarter of the size of the tiles in the example of
FIG. 8a). Because the tiles are smaller, each individual
16.times.16 tile involves less processing than an individual
32.times.32 tile, so the most costly 16.times.16 tiles are less
costly than the most costly 32.times.32 tiles, such that expensive
tiles being scheduled near the end of the render do not take as
long to process. The render is complete at time 808 with
16.times.16 tiles. It can be seen that time 808 is earlier than
time 804. This illustrates that in this example the total rendering
time for the render is shorter with 16.times.16 tiles compared to
with 32.times.32 tiles. The reduction to a 16.times.16 tile size
has three primary effects: 1) Work distribution becomes more
balanced across the processing cores (there are 1/4 the number of
pixels in each tile so the worst case slow tile is potentially 1/4
the length). 2) The render finishes sooner with the 16.times.16
tile size because up to 40% of the run time for the 32.times.32
tiles was being bottlenecked/wasted waiting on a small number of
slow running tiles. 3) Work in the render increases overall because
a smaller tile size is less efficient (primitives that were in only
one tile may now be present in 4 tiles as they cover the same
screen area). This can be seen in FIGS. 8a and 8b in that across
all of the processing cores, more processing is performed with
16.times.16 tiles than with 32.times.32 tiles (i.e. the area of
black lines in FIG. 8b is larger than the area of black lines in
FIG. 8a).
This third point means that smaller tiles may mean a larger amount
of processing for performing a render, which may increase the
processing power. This is a reason why it is not always beneficial
to keep reducing the tile size.
Inefficiencies occur where some processing cores are still
processing tiles when other processing cores are idle because they
have finished processing tiles assigned to them for the current
render. The greater the imbalance in workload towards the end of a
render the worse the efficiency/utilisation of the GPU cores, i.e.
the more time processing cores spend in an idle state. The problem
only appears at the end of a render because even though different
tiles may involve different amounts of work (i.e. the processing
costs may be different for different tiles) throughout most of the
render the basic scheduling policy used by the scheduling logic 316
still assigns work to the processing cores 314 for high utilisation
most of the time. In some examples, tiles can be subdivided into
smaller tiles near the end of a render, but not near the start of a
render because it is only at the end of a render that the problem
of larger tiles presents itself. Furthermore, this means that tiles
near the start of the render can have a larger tile size which can
be beneficial in terms of reducing the amount of processing
involved in processing an area of the rendering space. In other
words, the effective tile size can be changed towards the end of
renders where the load balancing is most beneficial without
incurring the general penalty of reducing the tile size across the
entire render.
Furthermore, it may be desirable to optionally apply the
subdivision of tiles to only those renders that exhibit tasks that
"stick out" substantially from the end of renders. A task "sticks
out" substantially from the end of a render if the task executes on
a particular processing core for a substantial amount of time
whilst other processing cores are idle because there are no more
tasks to process for the render. Whether a render exhibits such
tasks can be determined in several ways, e.g. based on a heuristic
using previous frame information, using cost indications generated
in any of the ways outlined above by the cost indication logic 320,
or based on a threshold number of tiles that are still to be
processed in the current render, where this threshold number may be
fixed or may be related to either the total number of processing
cores in the GPU or the number of idle processing cores in the GPU.
A decision on whether to subdivide tiles could be based on the
number of tiles (N) still to be scheduled and on the number of idle
cores (M).
FIG. 9 is a flow chart for a method involving the subdivision of
tiles. The geometry processing phase may be the same as in the
example described above with reference to FIG. 4. In other words,
steps S902 to S908 may be the same as steps S402 to S408 described
above, although in this example the use of cost indications for
tiles is optional, so steps S906 and S908 might not be performed in
some examples.
It is noted that the steps shown in the flow charts in FIGS. 4 and
9 may take place in a sequence, or some of the steps may be
performed simultaneously, i.e. in parallel. Furthermore, the steps
do not necessarily occur in the order shown in the Figures.
The scheduling logic 316 determines how the tiles are to be
assigned to the processing cores 314, i.e. how the tiles are to be
scheduled for rendering on the processing cores 314. According to
terminology used herein, a tile can be scheduled for rendering on a
processing core by assigning the tile to the processing core for
rendering. In step S910 the scheduling logic 316 determines which,
if any, of the tiles are to be subdivided for the current render.
For example, the scheduling logic 316 may determine which of the
tiles to subdivide for the current render based on information
relating to processing costs for corresponding tiles in a previous
render. If a tile in a previous render (e.g. the immediately
preceding frame in a sequence of frames being rendered) was
particularly expensive to process (e.g. because it includes many
primitives with an object type such as translucent or punch through
or because the processing of the tile involves the execution of
long shader programs) then the scheduling logic 316 may determine
that it would be beneficial to subdivide that tile for the current
render. As another example, if a tile was subdivided in a previous
render then the scheduling logic 316 may determine that the tile
should be subdivided in the current render also.
As another example, the scheduling logic 316 may determine which of
the tiles to subdivide for the current render based on a
relationship between the number of tiles that are still to be
rendered in the current render and a threshold number of tiles. In
this way the scheduling logic 316 may determine that tiles are to
be subdivided only if they are near to the end of the render. The
scheduling logic 316 may determine which of the tiles to subdivide
for the current render further based on the number of said
processing cores. For example, if there are more processing cores
then to be considered "near" to the end of the render may mean that
the threshold number of tiles is increased.
As another example, the scheduling logic 316 may determine which of
the tiles to subdivide for the current render based on the cost
indications for the tiles. For example, only expensive tiles might
be subdivided because these are the tiles that might cause a render
to be delayed whilst some processing cores are idle at the end of
the render. The cost indications may be determined by the cost
indication logic 320 as described above, and may be quantised as
described above.
The scheduling logic 316 may use a metric to determine which of the
tiles to subdivide for the current render. For example, the
scheduling logic 316 may subdivide tiles according to a metric
according to which a likelihood of a tile being subdivided is
increased if a cost indication that is associated with the tile is
increased. Alternatively, or additionally, the scheduling logic 316
may subdivide tiles according to a metric according to which a
likelihood of a tile being subdivided is increased if the number of
tiles of the current render that are still to be rendered is
decreased. Alternatively, or additionally, the scheduling logic 316
may subdivide tiles according to a metric according to which a
likelihood of a tile being subdivided is increased if the number of
processing cores is increased. The metric may take one or more
factors into account to allow the scheduling logic 316 to make a
decision as to which tiles to subdivide.
Step S910, to determine which tiles to subdivide, may be performed
up front in a pass which makes a decision for all tiles about which
will be processed and how. In some examples, the subdivision
process could be performed as a loop over each tile, such that for
each tile, the cost estimation values are read and a decision is
made about whether the tile should be subdivided or not, and then
the loop goes around to process the next tile.
In step S912 the scheduling logic 316 subdivides tiles which were
determined as being for subdivision in step S910. Tiles may be
subdivided into any suitable number of subtiles, e.g. tiles may be
subdivided into two halves or four quarters. For example a
32.times.32 tile may be subdivided into four 16.times.16 subtiles.
There is no requirement that the system be limited to modulo 2
values, so a tile may be subdivided into three subtiles as an
example. Similarly, there is no requirement that the subdivision
must subdivide the tile equally among the processing cores. For
example, a tile which includes 32.times.32 pixels may be divided
amongst five idle cores. If the tile was divided equally amongst
the cores, then that would result in 204.8 pixels per tile. In a
practical system, slightly different sized subtiles may be provided
to the different cores, e.g. such that each core receives a
(probably rectangular) block of pixels to process, rather than
render the same pixel in multiple subtiles. Although it may be
simpler to subdivide a tile into subtiles of equal size and shape,
there is no requirement that a tile is subdivided into subtiles of
equal size or equal shape.
FIG. 10a shows an example where a 32.times.32 tile 1002 is
subdivided to form four 16.times.16 subtiles denoted 1004, 1006,
1008 and 1010 in FIG. 10a. The separate subtiles can be assigned to
different processing cores for rendering. In this example, control
stream data may be determined for each of the subtiles 1004 to
1010. For example, separate control stream data may be determined
for each of the subtiles 1004 to 1010. Alternatively, the subtiles
1004 to 1010 may share the control stream data for the tile 1002
and then a mask (e.g. a 4-bit mask where there are four subtiles)
may be included for each primitive identified in the control stream
to indicate which of the four subtiles the primitive is valid in.
If a particular subtile (e.g. subtile A 1004) is assigned to a
particular processing core (e.g. processing core 314.sub.1) that
particular processing core (e.g. 314.sub.1) can receive the control
stream data for the particular subtile (e.g. 1004) in order to
render the particular subtile (e.g. 1004).
FIG. 10b shows another example of subdividing a 32.times.32 tile
1012. In this example, the subdivision process does not separate
the tile 1012 into smaller subtiles to be provided to processing
cores 314. Instead, the whole tile 1012 (i.e. the control stream
for the whole tile) is provided to the appropriate processing cores
along with a mask 1014 (which uses very little data, e.g. a 2-bit
mask in the example that the tile 1012 is subdivided into quarters)
which indicates which portion of the tile 1012 is valid for a
particular subtile. In a more flexible example, a mask for a
particular subtile may include a bit per pixel to indicate whether
each particular pixel is valid for the particular subtile. This
would allow more flexibility in the way in which a tile is
subdivided, but it means that the masks include more data (e.g. a
bit per pixel, rather than just two bits for the whole tile (which
may for example include 1024 pixels)). The hatched region of a mask
1014 represents an invalid region, whilst the unhatched region of a
mask 1014 represents a valid region of the tile 1012. In other
words, to modify the effective tile size a tile with a mask can be
submitted to indicate which pixels of the tile should be processed
by a given processing core. Extending this idea further we can
split the work of any tile across multiple processing cores by
submitting the same tile to `N` processing cores and providing
masks to indicate which pixels each processing core is responsible
for.
For example, the scheduling logic 316 may determine a plurality of
masks (1014.sub.A, 1014.sub.B, 1014.sub.C and 1014.sub.D) which
indicate valid regions of the tile 1012 for the respective
plurality of subtiles (A, B, C and D), wherein a particular
processing core is configured to receive data for a particular
subtile assigned to the particular processing core by receiving:
(i) data for a particular tile 1012 from which the particular
subtile is derived, and (ii) a respective mask for the particular
subtile. For example, subtile A may be assigned to processing core
314.sub.1 so the processing core 314.sub.1 will receive the control
stream for the tile 1012 and the mask 1014.sub.A which indicates
that the top left quadrant of the tile 1012 is valid for subtile A;
subtile B may be assigned to processing core 314.sub.2 so the
processing core 314.sub.2 will receive the control stream for the
tile 1012 and the mask 1014.sub.B which indicates that the top
right quadrant of the tile 1012 is valid for subtile B; subtile C
may be assigned to processing core 314.sub.3 so the processing core
314.sub.3 will receive the control stream for the tile 1012 and the
mask 1014.sub.C which indicates that the bottom left quadrant of
the tile 1012 is valid for subtile C; and subtile D may be assigned
to processing core 314.sub.4 so the processing core 314.sub.4 will
receive the control stream for the tile 1012 and the mask
1014.sub.0 which indicates that the bottom right quadrant of the
tile 1012 is valid for subtile D.
In step S914 tiles and subtiles are assigned to the processing
cores for rendering. This may be done in a similar to that
described above, e.g. by adding identifiers of the tiles or
subtiles to a rendering queue 318 for each of the processing cores
314. Tiles and/or subtiles may be assigned to a processing core
based on the workload queued for the processing core, e.g. when the
number of tiles and/or subtiles currently queued for a particular
processing core drops below a threshold number then one or more
tiles and/or subtiles can be assigned to the particular processing
core.
At least some of the subtiles derived from the same tile are
assigned to different processing cores for rendering. In other
words, subtiles derived from a tile are distributed to a plurality
of the processing cores 314 for rendering. For example, the
subtiles (e.g. subtiles A to D) derived from a particular tile
(e.g. tile 1012) are assigned to a plurality of the processing
cores (e.g. to processing cores 314.sub.1 to 314.sub.4), i.e. one
or more subtile (e.g. subtile A) derived from the particular tile
1012 is assigned to a first processing core (e.g. 314.sub.1) and
one or more subtile (e.g. subtile B) derived from the particular
tile 1012 is assigned to a second processing core (e.g. 314.sub.2),
where the first processing core (e.g. 314.sub.1) is a different
processing core to the second processing core (e.g. 314.sub.2).
Steps S916 and S918 are similar to steps S412 and S414 described
above. In step S916 the processing cores 314 render the assigned
tiles and assigned subtiles. The rendering is performed as
described above, e.g. by performing hidden surface removal and
texturing/shading on primitive fragments which are present within
the tiles or subtiles that are being rendered.
In step S918 the results of the render are stored, e.g. in the
memory 304.sub.2.
The method of subdividing tiles into subtiles and processing the
subtiles on different processing cores may be less efficient (e.g.
in terms of the amount of processing power used) than processing a
full tile on a single processing core because redundant data may be
read and processed for a tile on each of the multiple processing
cores which process subtiles derived from the same tile. However,
when the processing overhead is lower than the penalty of not load
balancing the system the net effect of subdividing some tiles (e.g.
subdividing expensive tiles near the end of a render) can be
considered to be an overall improvement for the system. This is why
even the extreme example shown in FIGS. 8a and 8b in which the tile
size for all tiles is quartered, the smaller tile size still comes
out faster than being bottlenecked by the slow tiles. The hybrid
subdivision method described herein where large tile sizes are used
for some tiles of a render (e.g. tiles near the start of a render
and/or tiles which have low cost indications) and smaller tile
sizes are used for other tiles (e.g. tiles near the end of a render
and/or tiles which have high cost indications) for example by
subdividing the larger tiles, results in a system that is faster
than the large tile size due to the load balancing improvements and
faster than the small tile size due to incurring the overhead only
for some of the tiles. Counterintuitively a source of inefficiency
has been added to a small part of the render (by reducing the tile
size for some tiles) to increase the efficiency of the system as a
whole.
As described above the scheduling logic 316 may assign the tiles to
the processing cores 314 in sets of one or more tiles. Furthermore,
as described above, each of the processing cores 314 may have
access to at least one cache for storing data (e.g. primitive data
or texture data) to be used for rendering primitives in a tile. The
scheduling logic 316 may assign subtiles derived from the same tile
to processing cores 314 which have access to the same cache. In
this way the data for the tile can be stored in the cache and used
by each of the processing cores, with appropriate masks 1014, to
render the different subtiles derived from that tile. This may
reduce the number of caches in which the data for the tile is
stored, which can improve the efficiency of the cache system.
Similarly, it may be desirable to be aware of the underlying cache
structure when subdividing and scheduling tiles for different
cores, e.g. to try to avoid scheduling sub-tiles derived from the
same tile such that they are processed on cores that use different
caches.
In the examples described above, tiles are subdivided into
subtiles. However, the same subdivision process can be used to
subdivide a set of tiles into smaller subsets of tiles. In general,
a set of one or more tiles can be subdivided into a plurality of
subunits, where this subdivision can either be a subdivision of a
tile into a plurality of subtiles (such that the "subunits" are
subtiles), or a subdivision of a set of tiles into a subset of one
or more tiles (such that the "subunits" are subsets of one or more
tiles). The number of tiles in the set of one or more tiles (prior
to any subdivision) is an implementation choice, e.g. in some
implementations single tiles are scheduled to processing cores
whereas in other implementations groups of tiles are scheduled to
processing cores.
As an example, sets of six tiles (e.g. 3.times.2 blocks of tiles)
may be submitted to a single processing core, and each set may have
a cost estimate (e.g. a quantised cost estimate) stored for it. The
individual tiles within the set may, or may not, have individual
cost estimates stored for them. The set of tiles may be subdivided
into smaller subsets of tiles (e.g. two subsets of three tiles
each, three subsets of two tiles each, or six subsets of one tile
each, to give just some examples) to be sent to a plurality of
available processing cores. For example, if the final set of tiles
submitted right at the end of the render is a 3.times.2 block of
tiles and the cost estimate for the set is very high (e.g. because
the pre-pass over all the tile costs has not been performed) then
it might be more efficient to subdivide the set of tiles into
multiple subsets of tiles (e.g. six single tiles) to be processed
by multiple cores (e.g. six cores) before considering whether to
further subdivide any of the tiles into subtiles.
There is a tradeoff of cache efficiency on the level of subdivision
that makes sense for each set of tiles, which may be dependent upon
the estimated cost of processing the set of tiles and the
underlying architectural cache layouts. For example, if four
processing cores are connected to the same cache then there may be
a relatively small penalty to splitting a set of tiles into four
subunits, but if a set of tiles is split up into greater than four
subunits (e.g. into six subunits) then there may be a relatively
large penalty as the set of tiles would then span into another
cache (i.e. data for the set of tiles would need to be written into
another cache).
The processing logic 313 may include a large number of the
processing cores 314, e.g. in some examples the processing logic
313 may include at least 100 processing cores. When the number of
tiles is not much greater than the number of processing cores it
becomes more likely that a tile will be scheduled "near the end" of
a render, and as such it becomes more likely that an expensive tile
may delay the render finish time. Therefore, with such a large
number of processing cores 314, it can become more beneficial to
subdivide the sets of one or more tiles, in particular the
expensive sets of one or more tiles.
In the subdivision methods described above, the scheduling logic
316 subdivides a tile into smaller subtiles. However, in other
examples, the tiling unit 308 could process some tiles with a
smaller size than other tiles in a render to achieve a similar
result in that some tiles can have a smaller effective tile size
than other tiles. For example, if tiles are to be assigned in a
raster scan order (i.e. not according to cost indications) then the
tiling unit will know which tiles are to be assigned to a
processing core near the end of a render (e.g. the bottom row(s) of
tiles of the rendering space), and as such it may be useful to
subdivide these tiles into smaller tiles such that control stream
data is determined in the tiling unit 308 for the subtiles near the
end of the render. In other examples, the tiling unit 308 may
decide which tiles are to be rendered with a smaller tile size
based on other factors. In this example, larger tiles are not
subdivided in the scheduling logic 316, and instead some tiles are
processed with a smaller size than other tiles in the geometry
processing phase and in the rendering phase, so in a sense these
tiles can be considered to be subdivided in the tiling phase.
In the examples described above the graphics processing system 300
comprises a plurality of processing cores 314. However, in some
examples, a graphics processing system may include just a single
processing core. The subdivision methods described above can be
useful for subdividing work which is all to be processed by a
single processing core. Therefore, the scheduling logic may assign
all of the subtiles derived from a tile to a single processing
core. This can be useful, e.g. if the processing core is able to
switch between processing different subtiles. For example, the
processing core may be able to store the state of a partially
processed subtile (e.g. the state of a depth buffer and a tag
buffer) while performing processing on a different one of the
subunits. In this way, the processing core may be considered to
have "multiple subtiles in flight". Having multiple subtiles in
flight may be more efficient (in some scenarios) than having a
single, larger tile in flight, because the parallelism of the
processing performed by the processing core can be improved. The
concept of having multiple tiles in flight in the rendering logic
is described in granted patent GB2520365, and this concept can be
extended to having multiple subtiles in flight. Similar advantages
can be achieved.
For example, a graphics processing system may be configured to
allow the processing of primitives to switch between primitives of
different subtiles before all of the primitives of a particular
subtile have finished being processed (thereby allowing the
graphics processing system to have "multiple subtiles in flight",
i.e. multiple subtiles for which the primitives are partially
processed at a given time). In order to achieve this the graphics
processing system may comprise a plurality of depth buffers for
storing depths of respective subtiles in flight and other memory
for storing other state data (e.g. a plurality of tag buffers for
storing tags of primitives for which depths are stored in
respective ones of the plurality of depth buffers) for a partially
processed subtile while a different subtile is being processed. If
the system processes the subtiles of a single full tile at a time
then the storage would not necessarily be increased (compared to
processing full tiles without subdivision) because the storage
(e.g. depth buffers and buffers for other state data) for the full
tile can be subdivided and redistributed to the subtiles. Splitting
the work within a single processing core by subdividing a tile into
subtiles means that a greater number of smaller passes are
performed, which can allow the work to flow more smoothly through
the processing core. For example, filling larger buffers (e.g. a
depth buffer and a tag buffer) for a whole tile with results of
depth testing could take long enough for shading engines to run out
of work (or begin to run out of work), thereby reducing the
efficiency of the shading engines. In contrast, results of depth
testing for subtiles may be written into smaller buffers
(corresponding to the subtiles) more quickly so it is less likely
that the shading engines will run out of work when the graphics
processing system processes subtiles rather than tiles.
Furthermore, dependencies within areas of a tile (e.g. overlapping
punch-through primitives) can cause stalls in the processing flow;
whereas subdividing a tile can allow subtiles to be processed in
parallel such that if one subtile stalls a different subtile can be
processed. This reduces the likelihood of the whole system
stalling.
Processing subtiles rather than tiles can slightly reduce the
processing efficiency (e.g. SIMD processing efficiency) because a
primitive spanning multiple subtiles will be processed as multiple
separate tasks which may run at different times, such that the task
occupancy may be reduced. However, this disadvantage can be
outweighed at the end of a render by the advantage of reducing the
total render time, e.g. for the last tile left running in a core.
Therefore, in some examples, the subdivision might only be used for
tiles near the end of a render.
In the description above, methods of determining the order in which
tiles are processed in the processing logic 313 based on cost
indications for the tiles are described, and methods of subdividing
sets of one or more tiles to be rendered are described. These
methods may be used separately or in combination. That is, each
method is useful in its own right and can reduce the render time
when implemented without the other method. Using both methods
together may reduce the render time compared to using just one of
the methods, particularly if there are a large number of processing
cores (e.g. 100 processing cores) implemented in the processing
logic 313. When using both methods together the more costly tiles
are processed first, and some of the tiles are subdivided if they
may delay the render finish time while some processing cores are
idle.
FIG. 11 shows a computer system in which the graphics processing
systems described herein may be implemented. The computer system
comprises a CPU 1102, a GPU 302, a memory 1106 and other devices
1108, such as a display 1110, speakers 1112 and a camera 1114. The
components of the computer system can communicate with each other
via a communications bus 1118. A store 1116 (corresponding to
memories 304.sub.1 and 304.sub.2) is implemented as part of the
memory 1106.
The graphics processing system 300 of FIG. 3 is shown as comprising
a number of functional blocks. This is schematic only and is not
intended to define a strict division between different logic
elements of such entities. Each functional block may be provided in
any suitable manner. It is to be understood that intermediate
values described herein as being formed by a graphics processing
unit need not be physically generated by the graphics processing
unit at any point and may merely represent logical values which
conveniently describe the processing performed by the graphics
processing unit between its input and output.
The graphics processing units described herein may be embodied in
hardware on an integrated circuit. The graphics processing units
described herein may be configured to perform any of the methods
described herein. Generally, any of the functions, methods,
techniques or components described above can be implemented in
software, firmware, hardware (e.g., fixed logic circuitry), or any
combination thereof. The terms "module," "functionality,"
"component", "element", "unit", "block" and "logic" may be used
herein to generally represent software, firmware, hardware, or any
combination thereof. In the case of a software implementation, the
module, functionality, component, element, unit, block or logic
represents program code that performs the specified tasks when
executed on a processor. The algorithms and methods described
herein could be performed by one or more processors executing code
that causes the processor(s) to perform the algorithms/methods.
Examples of a computer-readable storage medium include a
random-access memory (RAM), read-only memory (ROM), an optical
disc, flash memory, hard disk memory, and other memory devices that
may use magnetic, optical, and other techniques to store
instructions or other data and that can be accessed by a
machine.
The terms computer program code and computer readable instructions
as used herein refer to any kind of executable code for processors,
including code expressed in a machine language, an interpreted
language or a scripting language. Executable code includes binary
code, machine code, bytecode, code defining an integrated circuit
(such as a hardware description language or netlist), and code
expressed in a programming language code such as C, Java or OpenCL.
Executable code may be, for example, any kind of software,
firmware, script, module or library which, when suitably executed,
processed, interpreted, compiled, executed at a virtual machine or
other software environment, cause a processor of the computer
system at which the executable code is supported to perform the
tasks specified by the code.
A processor, computer, or computer system may be any kind of
device, machine or dedicated circuit, or collection or portion
thereof, with processing capability such that it can execute
instructions. A processor may be any kind of general purpose or
dedicated processor, such as a CPU, GPU, System-on-chip, state
machine, media processor, an application-specific integrated
circuit (ASIC), a programmable logic array, a field-programmable
gate array (FPGA), or the like. A computer or computer system may
comprise one or more processors.
It is also intended to encompass software which defines a
configuration of hardware as described herein, such as HDL
(hardware description language) software, as is used for designing
integrated circuits, or for configuring programmable chips, to
carry out desired functions. That is, there may be provided a
computer readable storage medium having encoded thereon computer
readable program code in the form of an integrated circuit
definition dataset that when processed in an integrated circuit
manufacturing system configures the system to manufacture a
graphics processing unit configured to perform any of the methods
described herein, or to manufacture a graphics processing unit
comprising any apparatus described herein. An integrated circuit
definition dataset may be, for example, an integrated circuit
description.
Therefore, there may be provided a method of manufacturing, at an
integrated circuit manufacturing system, a graphics processing unit
as described herein. Furthermore, there may be provided an
integrated circuit definition dataset that, when processed in an
integrated circuit manufacturing system, causes the method of
manufacturing a graphics processing unit to be performed.
An integrated circuit definition dataset may be in the form of
computer code, for example as a netlist, code for configuring a
programmable chip, as a hardware description language defining an
integrated circuit at any level, including as register transfer
level (RTL) code, as high-level circuit representations such as
Verilog or VHDL, and as low-level circuit representations such as
OASIS.RTM. and GDSII. Higher level representations which logically
define an integrated circuit (such as RTL) may be processed at a
computer system configured for generating a manufacturing
definition of an integrated circuit in the context of a software
environment comprising definitions of circuit elements and rules
for combining those elements in order to generate the manufacturing
definition of an integrated circuit so defined by the
representation. As is typically the case with software executing at
a computer system so as to define a machine, one or more
intermediate user steps (e.g. providing commands, variables etc.)
may be required in order for a computer system configured for
generating a manufacturing definition of an integrated circuit to
execute code defining an integrated circuit so as to generate the
manufacturing definition of that integrated circuit.
An example of processing an integrated circuit definition dataset
at an integrated circuit manufacturing system so as to configure
the system to manufacture a graphics processing unit will now be
described with respect to FIG. 12.
FIG. 12 shows an example of an integrated circuit (IC)
manufacturing system 1202 which is configured to manufacture a
graphics processing unit as described in any of the examples
herein. In particular, the IC manufacturing system 1202 comprises a
layout processing system 1204 and an integrated circuit generation
system 1206. The IC manufacturing system 1202 is configured to
receive an IC definition dataset (e.g. defining a graphics
processing unit as described in any of the examples herein),
process the IC definition dataset, and generate an IC according to
the IC definition dataset (e.g. which embodies a graphics
processing unit as described in any of the examples herein). The
processing of the IC definition dataset configures the IC
manufacturing system 1002 to manufacture an integrated circuit
embodying a graphics processing unit as described in any of the
examples herein.
The layout processing system 1204 is configured to receive and
process the IC definition dataset to determine a circuit layout.
Methods of determining a circuit layout from an IC definition
dataset are known in the art, and for example may involve
synthesising RTL code to determine a gate level representation of a
circuit to be generated, e.g. in terms of logical components (e.g.
NAND, NOR, AND, OR, MUX and FLIP-FLOP components). A circuit layout
can be determined from the gate level representation of the circuit
by determining positional information for the logical components.
This may be done automatically or with user involvement in order to
optimise the circuit layout. When the layout processing system 1204
has determined the circuit layout it may output a circuit layout
definition to the IC generation system 1206. A circuit layout
definition may be, for example, a circuit layout description.
The IC generation system 1206 generates an IC according to the
circuit layout definition, as is known in the art. For example, the
IC generation system 1206 may implement a semiconductor device
fabrication process to generate the IC, which may involve a
multiple-step sequence of photo lithographic and chemical
processing steps during which electronic circuits are gradually
created on a wafer made of semiconducting material. The circuit
layout definition may be in the form of a mask which can be used in
a lithographic process for generating an IC according to the
circuit definition. Alternatively, the circuit layout definition
provided to the IC generation system 1206 may be in the form of
computer-readable code which the IC generation system 1206 can use
to form a suitable mask for use in generating an IC.
The different processes performed by the IC manufacturing system
1202 may be implemented all in one location, e.g. by one party.
Alternatively, the IC manufacturing system 1202 may be a
distributed system such that some of the processes may be performed
at different locations, and may be performed by different parties.
For example, some of the stages of: (i) synthesising RTL code
representing the IC definition dataset to form a gate level
representation of a circuit to be generated, (ii) generating a
circuit layout based on the gate level representation, (iii)
forming a mask in accordance with the circuit layout, and (iv)
fabricating an integrated circuit using the mask, may be performed
in different locations and/or by different parties.
In other examples, processing of the integrated circuit definition
dataset at an integrated circuit manufacturing system may configure
the system to manufacture a graphics processing unit without the IC
definition dataset being processed so as to determine a circuit
layout. For instance, an integrated circuit definition dataset may
define the configuration of a reconfigurable processor, such as an
FPGA, and the processing of that dataset may configure an IC
manufacturing system to generate a reconfigurable processor having
that defined configuration (e.g. by loading configuration data to
the FPGA).
In some embodiments, an integrated circuit manufacturing definition
dataset, when processed in an integrated circuit manufacturing
system, may cause an integrated circuit manufacturing system to
generate a device as described herein. For example, the
configuration of an integrated circuit manufacturing system in the
manner described above with respect to FIG. 12 by an integrated
circuit manufacturing definition dataset may cause a device as
described herein to be manufactured.
In some examples, an integrated circuit definition dataset could
include software which runs on hardware defined at the dataset or
in combination with hardware defined at the dataset. In the example
shown in FIG. 12, the IC generation system may further be
configured by an integrated circuit definition dataset to, on
manufacturing an integrated circuit, load firmware onto that
integrated circuit in accordance with program code defined at the
integrated circuit definition dataset or otherwise provide program
code with the integrated circuit for use with the integrated
circuit.
The implementation of concepts set forth in this application in
devices, apparatus, modules, and/or systems (as well as in methods
implemented herein) may give rise to performance improvements when
compared with known implementations. The performance improvements
may include one or more of increased computational performance,
reduced latency, increased throughput, and/or reduced power
consumption. During manufacture of such devices, apparatus,
modules, and systems (e.g. in integrated circuits) performance
improvements can be traded-off against the physical implementation,
thereby improving the method of manufacture. For example, a
performance improvement may be traded against layout area, thereby
matching the performance of a known implementation but using less
silicon. This may be done, for example, by reusing functional
blocks in a serialised fashion or sharing functional blocks between
elements of the devices, apparatus, modules and/or systems.
Conversely, concepts set forth in this application that give rise
to improvements in the physical implementation of the devices,
apparatus, modules, and systems (such as reduced silicon area) may
be traded for improved performance. This may be done, for example,
by manufacturing multiple instances of a module within a predefined
area budget.
The applicant hereby discloses in isolation each individual feature
described herein and any combination of two or more such features,
to the extent that such features or combinations are capable of
being carried out based on the present specification as a whole in
the light of the common general knowledge of a person skilled in
the art, irrespective of whether such features or combinations of
features solve any problems disclosed herein. In view of the
foregoing description it will be evident to a person skilled in the
art that various modifications may be made within the scope of the
invention.
* * * * *